Consumer health startup Teal Health raised $22 million last month to get more women screening for cervical cancer at home with their Teal Wand.

Starting in the 1940s, cervical cancer screening in the US meant a Pap smear: women booking a visit, taking time off, undressing, and lying back, feet in stirrups, while a clinician inserted a speculum and collected cells from the cervix. HPV testing was later added as the preferred test, but the visit and the speculum stayed.

80+ years later, in May 2024 the FDA let women swab themselves, but only in a health care setting. No speculum, but you still had to take time off work, go to the clinic, etc.

Then, on May 9, 2025, the FDA authorized the Teal Wand, which lets a woman collect her own vaginal sample at home and mail it to a lab for HPV testing.

No doc, no clinic, no time off.

Now insurers have to pay for DIY

In January, HRSA updated its guidelines for women's preventive care to include self-collected HPV testing. Under the Affordable Care Act, that means most private insurance plans have to pay for it, with no patient cost, starting in 2027 — including any follow-up tests, like colposcopy, needed after a positive result.

Teal already works with Aetna, Cigna, Blue Cross Blue Shield, and UnitedHealthcare. Those plans pay for the prescription, the lab, and the results, but not the kit or the mailing, so women pay $99 with insurance or $249 without. Next year women on those plans and others will pay nothing.

Does it work?

Yes.

In Teal's own study, published in JAMA Network Open, about 600 women swabbed themselves and then had a clinician collect a sample too. The self-swab caught 95% of the HPV infections the clinician's sample found. The test looks for HPV, not precancer (which needs a biopsy). But the study did those biopsies, and among the women whose biopsies did show precancer, the self-swab had tested HPV-positive 96% of the time, same as the clinician’s sample.

Older research had similar results. A large 2018 review found self-swabs work about as well as clinician samples. Mailing kits to women's homes more than doubled the number who got screened, but making women request a kit didn't help much.

Teal today works the second way: a woman orders her own Wand — pull, not push, so that alone shouldn’t increase screening numbers. But Teal says the Series A will fund partnerships with insurers, health systems, and employers — who could mail kits to everyone and bump up those numbers.

What still needs the clinic

The lab tests the swab for high-risk HPV, the virus behind nearly all cervical cancer. If it's negative, the woman should screen again in three years. If it's positive, she goes back to the clinic for more testing, often colposcopy, where a clinician magnifies the cervix and biopsies anything suspicious. So followup is always in-clinic.

Also: the Teal Wand is only for average-risk women aged 30–65. Women under 30 and high-risk women also are always screened in-clinic.

Bottom line: Teal is classic Diagnosis Migration to the consumer sector, but follow-up diagnosis and treatment remain in the clinic.

Who gains

The women who gain most are those who haven't been screened at all, or not recently, and this is an increasing number of women. The share of US women 30–65 overdue for screening rose from 14.5% in 2005 to 21.1% in 2019, per a 2022 JAMA Network Open study. Teal says 59% of its users are underscreened, and its CEO told Fierce Healthcare that it finds cancer at about five times the rate expected in general screening. Those are from the company itself, but I think that’s about what you'd expect if the test can reach the women who aren’t getting traditional screening.

Will she miss the annual visit?

If women using Teal skip visits, which they very likely will, what do they lose? Teal's telehealth visit covers eligibility, the prescription, results, and coordination of follow-up. It doesn't do a pelvic exam, take a blood pressure, give vaccines, or handle contraception, STI testing, or mammogram referrals (at least not that it mentions on its site).

First: routine pelvic exams have no strong evidence to support it: USPSTF found the evidence insufficient to recommend for or against it, and the American College of Physicians recommends against it in asymptomatic women.

What about the rest of what's done at that annual visit? HRSA's own evidence update found no studies evaluating the effectiveness of well-woman visits (I should be shocked, but somehow I’m not). The closest evidence, a 2021 JAMA review, found no reduction in mortality or cardiovascular events, but better uptake of preventive services, more detection of conditions like hypertension and depression, and better blood pressure and cholesterol control.

Of course that was before hypertension-tracking wearables, DTC lab orders, and AI health coaches on our phones, so solve for the equilibrium, as the Marginal Revolution guys always say.

Who loses

That’s easy: gynecologists, plus the family doctors, internists, and NPs who do much of this screening. Good for their patients, but their offices will have fewer screening visits and fewer collection fees. Medicare's 2026 fee schedule pays $46.09 to obtain a screening Pap specimen (Q0091), plus $39.75 for a screening pelvic and breast exam (G0101), nationally, before local adjustment. Most women 30–65 have commercial insurance, which usually pays more. So that is on the line.

One offset, maybe: more women screened means more positives, and every positive goes back to a clinician for follow-up and colposcopy.

Another loser: the healthcare system employing those clinicians, if there is one. As I wrote in When Medical Testing Becomes Free, once the test is cheap or free, value shifts to whoever owns the relationship with the patient.

Which in this case will be Teal.

HealthCCSng, an awkward name even by medical-software standards, is an AI system developed by Zebra Medical Vision, now part of Nanox.AI. It analyzes routine chest CT scans and estimates coronary artery calcium, a marker of coronary artery disease. The CT may have been done to look for lung cancer, pneumonia, or something else but if it includes the heart, the software can use the existing images to run an additional cardiovascular test.

The FDA summary reports validation on 447 CT scans from two healthcare institutions, using three radiologists' majority opinion as the gold standard. We also know that 245 days elapsed between the FDA submission and clearance.

We can't tell from public records what HealthCCSng cost to develop and validate. Presumably, it wasn't cheap, but once the software has been validated the cost to run it on a scan is negligible: no scanner, no staff, no appointment, no radiation. It just looks at the scan that was already done.

As medical testing becomes digital, and costs plummet, its economics increasingly favor companies well-versed in digital business models — just as happened with phone calls, photography, recorded music, travel reservations, etc ad infinitum. As that happens, Big Tech and tech startups will increasingly dominate the development of new health tests and the interpretation of those and traditional medical tests and scans. Healthcare providers will lose revenue from interpretation as that work moves to consumer technology services. Laboratories and imaging centers will also lose revenue as new digital tests replace some of the conventional tests they sell.

When the test stops being the product

Email is a good example of how digitization changes economics. In 1983, MCI Mail charged $1 per message. In 1996, AOL charged by the hour, then moved to subscriptions. In 2004, Google launched Gmail as a free service to anyone with an internet connection.

None of those systems was free to build or operate. But as the marginal cost of sending another message fell, email providers stopped charging for individual messages. They sold network access, subscriptions, advertising, storage, hardware, business services, and a continuing relationship with the user. Today nobody ever considers the cost of an email before hitting send — or, similarly, the cost of a "long-distance call" (has anyone under the age of 30 even heard that phrase?).

Medical testing has three main cost elements: development, acquisition, and interpretation — and software can drive down all three. AI may reduce development costs by allowing companies to adapt models and infrastructure built for other purposes. When software runs on an existing scan, ECG, laboratory history, or sensor stream, the new test has no acquisition cost of its own, so that drops to zero. AI can also automate interpretation, reducing or eliminating the clinician time and the cost required for each additional analysis. As that cost trends towards zero, charging for each interpretation will start to look as strange as charging for each email.

What companies sell instead

A test that costs very little to run can be bundled with hardware, included in an AI subscription, given away to keep users inside an ecosystem, or used to sell another service. A company could charge an insurer for preventing admissions, take a share of savings, or get paid for achieving a health outcome.

Apple mostly wants to sell more hardware. It added hypertension notifications to watches already on users' wrists, using optical-sensor data already being collected. The new test arrived as software and costs the user nothing beyond the phone and watch (by now, smartphone users take it for granted that more features will be added by free software updates). Apple Health's AI interpretation adds more value still.

Google offers tiered pricing. Its free Google Health app combines health data from several sources; an additional $9.99 a month buys “Google Health Premium,” with coaching, readiness, and other features similar to Apple's.

OpenAI makes health interpretation part of a general AI service. ChatGPT Health can combine medical records and Apple Health data to explain laboratory results and identify patterns over time. Interpretation becomes a feature of the consumer's existing relationship with an AI service.

Another model would connect patients to care. Zocdoc already charges providers for new-patient bookings and sponsored placement. A tech company could offer free testing and/or interpretation and earn money from advertising and specialist bookings (subject to the Anti-Kickback Statute, which restricts payments for referrals involving federally covered care). A company paid for bookings would have a reason to turn findings into appointments, helping people obtain needed care but also potentially encouraging unnecessary follow-up.

The next king of testing will be a Big Tech company

Traditional healthcare has controlled which tests are ordered, where the resulting data are stored, and which interpretations reach the patient. Laboratories produce measurements, imaging centers produce images, and clinicians interpret the results.

Digitization can shift power from the company producing the underlying material to the company making a service from it. Twenty-odd years ago, musicians and labels still made the music, but Apple's iTunes Music Store became the dominant marketplace and helped sell highly profitable iPods. Apple captured the customer relationship and much of the value created by digitization.

Medical testing will follow the same path. Whoever has permission to use the data and the resources to develop new digital tests (i.e. new software) will own the next generation of medical testing.

Big Tech can disrupt medical testing in three ways:

1. Interpret existing tests. Consumer software like ChatGPT Health can interpret lab results and scans in the context of a person's medical history and also their everyday health data. That moves work and revenue away from the clinicians and healthcare organizations that currently sell interpretation.

2. Develop new tests for existing medical data. Software can extract additional signals from scans, ECGs, and digitized data from biological samples (as with HealthCCSng). The physical test still supplies the data, but a tech company can develop and provide the additional test, including for data collected years earlier, and capture the value of that additional test.

3. Replace conventional tests with software running on consumer devices. Apple's AirPods hearing test plays tones and uses your responses to produce an audiogram at home. In a 201-person validation study, its results closely matched audiologist-administered testing. You can obtain that hearing measurement without a clinic visit. Of course, this makes Apple hardware more valuable while displacing a paid test. The next step is algorithms that use everyday data from wearables, phones, rings, scales, and other consumer devices (i.e. not a specific hearing test the user has to stop and perform) to replace some lab tests or scans.

Competing with free

These activities create a competitive problem for the health systems, imaging vendors, and diagnostic-software companies that currently control how results are interpreted and delivered. They may hope to sell a new software test at a high margin — but a big tech company may be able to include a similar interpretation in a subscription the patient already pays for, or give it away. The tech giants can also quickly update their models across millions of users, while a hospital may wait months or years for its imaging vendor, budget cycle, regulatory review, and IT department to agree on an upgrade.

Laboratories and imaging centers will continue supplying data for tests that require specimens or new scans. New test development and interpretation will increasingly happen on platforms controlled by Big Tech, either through algorithms developed by Apple, Google, OpenAI, and their peers or through smaller diagnostic-software companies selling services in their app stores. Existing diagnostic companies may need those platforms to reach patients. Either way, Big Tech controls the customer relationship and takes a share of the value.

The bottom line: companies whose revenue depends on selling tests or interpretations will have to compete with companies that can profit by giving them away.

The next ten years

The postal service still delivers physical letters to physical mailboxes, and healthcare still uses fax. Hospitals will continue ordering and interpreting tests, trying to charge familiar prices even as software reduces their costs. For a shrinking share of the population, that will remain the only way medical testing happens.

Increasingly, though, people will order their own laboratory tests, obtain copies of their scans, and ask their phones and computers to interpret the results. More tests will come from Big Tech and from startups selling through the app stores.

The cost of testing will surely decrease as more becomes digital, but that may not reduce the costs of healthcare, as more findings may lead to more visits, procedures, and treatment — though more of the treatments may also be provided by tech companies (see today's announcement by Nolla Health of their pilot in Utah to diagnose, prescribe, and followup patients with acne — with no human clinician involved).

For traditional labs and healthcare providers, the question will be: how does a business built on selling tests or interpretations compete with a company that profits by giving them away?

First of two posts on medical (well, health-related) testing. Next week: what happens when the cost of running one more software-based test heads toward zero — and why cheaper testing may not mean cheaper healthcare.

“It’s about damn time.” — D. Griffin Jones on the new Siri update

My iPhone just automatically updated itself to iOS 27, and all of a sudden, at no extra cost, it has capabilities it didn’t have when I bought it. Chief among them: no kidding, Siri now actually answers my questions intelligently.

I didn’t buy another phone, replace any parts, or take it anywhere for an upgrade. New software made the device I already owned more useful. It’s amazing, but at this point we’ve become so accustomed to this that we barely stop to think about it.

This is already happening in health. I wrote about it in “Hidden Connections,” when Apple added hypertension notifications to watches people already owned. New software gave existing sensors another medical job.

Here I want to follow that process into the clinic. What happens when software finds something in an ordinary test that nobody ordered it to look for — and a doctor acts on the finding? We have some revealing examples, including one in which the scans had been sitting in the archive for years.

One test, several answers

As I discussed in my earlier post, this is happening across different kinds of tests. In one example that made the news, researchers at Google used AI to estimate cardiovascular risk from retinal photographs for the first time.

Mammograms, too. The arterial calcium visible in them — radiologists see it, but it’s not the reason for the scan — can be measured by AI to help estimate future cardiovascular risk. An Emory and Mayo team demonstrated this in a study of more than 123,000 women.

The question increasingly becomes what we do with that additional information.

Liver calling the heart

Now consider the ECG: a recording of the heart’s electrical activity that we’ve been doing for about 100 years. We know that liver cirrhosis affects the heart as well as the liver, and Mayo researchers showed those heart changes could be identified on an ECG. The test has always contained that information, but we just hadn’t been reading it.

In a follow-up clinical trial published last December, more than 15,000 patients received ECGs as part of routine care. Clinicians caring for roughly half of them received an alert when the AI identified a higher risk of advanced liver disease. Within six months, advanced fibrosis or cirrhosis had been newly diagnosed in 1% of that group, versus 0.5% with usual care.

Patients still needed follow-up testing to establish the diagnosis. But the reason to do liver testing came from existing ECGs, analyzed by new software. Nobody had to invent, or buy, a better ECG machine.

There’s a name for this: opportunistic screening — pulling additional clinically useful information out of a test ordered for some other reason. In this case, the software’s finding was presented to a clinician, prompted further investigation, and led to additional diagnoses and, presumably, treatment when required.

The scan was already in the archive

Stanford researchers decided to look at chest CTs that had already been performed for a variety of reasons. They ran AI over the images to identify coronary calcium, had radiologists confirm it, and randomly assigned patients to notification or usual care. The notifications went to patients and to their clinicians.

Six months later, because of the new finding in the old CT, 51% in the notification group had been prescribed a statin, versus 7% with usual care.

Think about this: the radiology archive had useful information, information that might add years to patients’ lives — but no one was looking. Why? Probably because even though we have known for awhile that the calcium signal was there, radiologists don’t have the time to add one more review to every chest CT they look at. But software has that time.

Not to say that there’s no potential downside: every new flag could trigger another test, procedure, diagnosis, or treatment. Overdiagnosis doesn’t become harmless because it’s easy, or cheap.

But the cirrhosis and Stanford studies show that these software changes can change what gets diagnosed or prescribed. Whether the additional diagnoses and prescriptions help people live longer or better is the further question we need to answer, in each case.

When the additional test is electrons

In 2011, Marc Andreessen wrote in The Wall Street Journal that “software is eating the world.” It was true for travel agents in 2011, and it’s now becoming true for medicine. As one part of that, we can increasingly use software to apply additional tests to information we already have without taking another specimen, or exposing the patient to more radiation, or scheduling another appointment.

Developing and validating that software costs money, and following up on its findings can run up the medical bill. But applying it to one more existing scan is a very different proposition from having to bring the patient in again, take another image, or draw more blood. That’s what makes these software developments so consequential.

This process isn’t limited to the hospital or clinic. I started this post talking about Siri updates, but my phone and my watch now offer fall detection, walking-steadiness monitoring, hypertension alerts, and many other useful health features, including screening tests.

The next question is what happens when the cost of running one more software-based test approaches zero. A test can be bundled with a phone, included in an AI subscription, or run by a company that never collected the original data. Will patients pay less? Who will interpret the results — and who will own the relationship with the patient — when medical testing becomes software?

I’ll write about that shortly.

That's a pretty definitive "never": no AI making management decisions. Ever.

The problem is that, as I point out in the letter, that “never” ended in 2018, when the FDA approved LumineticsCore's no-doc-in-the-loop system for diagnosing diabetic retinopathy. With LumineticsCore bearing responsibility for its AI’s medical judgment and carrying appropriate malpractice insurance.

Same for Doctronic's current prescription renewal pilot in Utah: no doc for most decisions means Doctronic had to get malpractice insurance.

Unfortunately, as Rao and Succi’s article suggests, much of the medical community still has not registered these developments — including the one from 2018 — and keeps treating AI accountability as a problem with no solution.

The solution for unaccountable practice of medicine is there, we just need to fully implement it — exactly as we do with human doctors.

Read the letter at STAT.

Classic Demand Elimination.

And over the study period the decline accelerated: roughly 14% in 2023, then 23% in 2024. Surgical volumes dropped below 200,000 procedures in 2024 for the first time since 2020.

A Wells Fargo Industry Insights report released last week points out the economic consequences, something my readers will already be familiar with. For years hospitals have been building service lines and revenue around all the downstream consequences of obesity: bariatric surgery, cardiology, orthopedics, sleep medicine, liver disease, diabetes, et al. GLP-1s reduce the obesity and should reduce all those downstream effects. (Fierce Healthcare coverage of the report).

Aside from this recent JAMA Surgery piece, we've seen other data confirming what's happening and what's to come. As one example, the SELECT trial showed a 20% reduction in major adverse cardiovascular events with semaglutide in overweight or obese adults without diabetes. There are also promising results in treating alcohol abuse.

One key insight from Calzaretta that may be overlooked: "more than 90% of the study population received neither form of treatment [i.e. surgery or GLP-1s] during the study period.

Translation: this is just getting started.

Plenty of near-term friction as pharmaceutical companies, employers, insurers and healthcare try to maneuver around this accelerating change. Drug spend is already lifting employer and payer costs, causing some large employers to consider dropping coverage for weight-loss drugs for 2027.

This, of course would be a colossal mistake from both the public health and the economic perspective: huge cost reductions will accrue as these meds reach more and more obese people. Employers just need to figure out how to finance that expansion, so they can reach the promised land of cost reduction in the future. More on this coming soon.

REMI monitors brain activity and identifies seizures. The standard procedure for that today means a stay in an epilepsy monitoring unit (EMU) — and only about 4% of U.S. hospitals have such a specialized center. The patient is wired to a full cap of scalp electrodes and lays in a hospital bed for 3 days, while simultaneous video is recorded. If a seizure happens outside that window, it’s invisible to the system unless the patient reports it.

REMI at home is worse than monitoring in the clinic on a few measures:

• lower resolution

• no video

• no clinician on hand during a seizure

But like many potentially disruptive products, it is much, much cheaper (no hospital admission) and much more convenient.

Even from the medical perspective, it has the huge advantage of monitoring for weeks, rather than three days, thereby possibly catching many more seizure episodes. And in a validation study, REMI’s system correctly identified all the seizures that the standard system spotted.

Bottom Line

About 9.2 million Americans have seizures each year, but only 4% of hospitals have EMUs. With REMI, the doctor prescribes the device, everything else happens at home (actually, I guess even the prescription could happen at home some of the time).

Classic Monitoring Migration.

From the hospital perspective, this is more bad news and revenue loss. From the epileptic patient perspective, it’s great: cost drops, convenience goes up, and extended monitoring might reach everyone who needs it, not only those who can get to an EMU.

And let’s not forget the potential increase in understanding of seizures that may come from a growing database of month-long EEG recordings.

The study, published in JAMA on July 6, documents what prescribing now requires: at 92% of sites, a self-reported questionnaire and an upper-body photo. The sites required no physician conversation, no records, no labs, no identity check. Four sites denied the application — two wanted blood work, one realized the same patient had already received a prescription elsewhere, one noticed the photo didn't match the reported weight.

From the patient side this is more like online shopping than traditional healthcare: answer some questions, pay, and the product arrives. That may not officially count as OTC, but it's indistinguishable in practice.

So what’s the harm

The Yale study describes web forms reviewed by licensed clinicians signing prescriptions at a pace — five minutes, in two cases — that suggests that not a lot of thought was applied. Millions of these prescriptions have gone out through this super-brief screening over two to three years. If it were truly dangerous, I suspect the harm would be broadly visible by now.

But ... it mostly isn't. GLP-1 adverse-event reports and poison-control calls have climbed sharply, but use is climbing faster, and the systems that tally the events count totals, not rates. And those events are mostly minor — nausea and vomiting — while serious outcomes stay a small fraction.

Dosing errors are probably the one adverse event worth paying attention to, and maybe the best argument against this loosely-regulated access.

Shut it down? Or open it up?

F. Perry Wilson, a Yale physician and director of Yale's Clinical and Translational Research Accelerator, covered the same study for Medscape and boiled it down to two possibilities: enforce the rules and shut down the pill mills, or just make the stuff OTC.

The study authors leaned on his first option, but Wilson seems open to both. I think that's good, because ratcheting up enforcement of the prescription requirement may solve the safety issue, but it leaves legitimate demand — a LOT of it — unfilled. Every day that we spend auditing prescribers and slowing the rate of use means another day that people suffer from obesity and move towards diabetes and all the other downstream complications.

So is there a safe way to say "yes" to everyone asking for these meds (who qualifies for them) instead of saying "no" (or "wait 6 months")?

The OTC case

GLP-1s have a 20-year record: exenatide since 2005, liraglutide since 2010, semaglutide since 2017. Nausea and GI effects dominate, pancreatitis is rare, the boxed thyroid warning comes from rat studies — we haven't seen that in people over the 20 years of use.

We have to weigh those minimal side effects against untreated obesity and its consequences: diabetes, heart disease, cancer, joint replacement. Every month access stays restricted, more people develop disease they could have avoided.

In addition, we have much more dangerous drugs on the OTC list, and have had for some time: insulin, for example, which can kill at the wrong dose, is OTC in most states. For GLP-1s, pregnancy is the most cited contraindication — although concerns deriving from animal data have not been borne out in seven cohorts covering more than 40,000 exposed pregnancies. Besides which, lingering concerns could be alleviated with the addition of a $1 pregnancy test in the box.

And the FDA now has a pathway for OTC that may require some screening: ACNU ("additional condition for nonprescription use"), finalized in December 2024. ACNU lets FDA approve a drug for sale without a prescription, using an app or other type of questionnaire to screen the buyer. No company has used it yet; GLP-1s would be a good candidate for first use.

If that happens, a GLP-1 application would look much like what the telehealth sites already do, except the screening would have to work, and it wouldn't cost $149 a month.

Where this ends up

We have precedent, a mild side-effect profile, and a new FDA pathway that would work to scale up population use of these beneficial drugs. The obstacles are the DTC providers, who don't want a free version of the questionnaire they're charging $149 for, and probably the drug companies themselves, who recognize that insurance coverage helps subsidize higher prices. And the drug companies are the ones who would have to file: ACNU requires an application, and in practice only the manufacturer has the data to build one.

My guess, though, is that the drug companies already see where GLP-1 prices are heading. Both Novo and Lilly now sell direct to patients outside insurance, and Novo advertised the starting dose of the Wegovy pill at $149 a month cash. They'll choose less profit per unit and a much larger population on the medication. That's better for patients and for population health, worse for the DTC companies, and — like anything that reduces obesity — a revenue killer for traditional healthcare.

A few days ago (July 16), Signal Ring (confusingly the name of the company and its only product) announced a $399 optical ring that reads blood pressure from your finger, continuously, with no cuff and no calibration — "the world's first calibration-free blood pressure wearable," per the company. It is supposed to ship in October.

This gave me a good opportunity to have a closer look at wearable BP technology, something I've been meaning to do for some time.

Monitoring of heart rate, oxygen, ECG, and sleep moved onto our wrists and fingers years ago, but blood pressure has taken longer. One approach has been to build a tiny cuff into the device, and measure pressure directly. A theoretically simpler approach, made possible by advances in AI, is to use an optical sensor to measure different characteristics of your pulse, which can then be correlated to BP. The problem has been to do so consistently, under a variety of scenarios (rising or falling BP, moving or staying still, etc), and most developers have settled on calibrating and recalibrating with a cuff (the gold standard) periodically.

Signal Ring, again, claims that it has overcome the problems.

What's on the market

The Huawei Watch D and D2 (not currently available in the US) are the tiny-cuff approach: the strap inflates and measures pressure directly, so the numbers are accurate. Both passed ISO 81060-2 — the standard test for a cuff, which checks its readings against a trained observer with a stethoscope across a range of patients — in published studies, including one independent trial. But readings are periodic, not continuous, and it isn't FDA-cleared.

Samsung's Galaxy Watch and the Hilo band use the optical + cuff-calibration approach. Samsung released this feature in the US last March, as a wellness feature it says isn't for medical use. This is a common way around FDA review (see FDA's general-wellness policy), and so at this point it's neither FDA-cleared nor validated. Hilo went the other way, applying for FDA clearance as a medical device. In July 2025 it became the first cuffless monitor to get that clearance, and it is due to launch in the US this year.

Samsung's Galaxy Watch and the Hilo band use the optical, cuff-calibration approach. Samsung reached the US on March 31, 2026, as a wellness feature it says isn't for medical use. This is a common way around FDA review (see FDA's general-wellness policy), and so at this point it's neither FDA-cleared nor validated. Whoop's top band does the same, reporting a daily BP range built on an initial cuff calibration and periodic manual checks — accuracy some users have questioned. Hilo went the other way, applying for FDA clearance as a medical device. In July 2025 it became the first cuffless monitor to get that clearance, and it is due to launch in the US this year.

Apple's watch doesn't measure blood pressure at all. Since September 2025 it reads the optical sensor over 30-day stretches and flags signs of chronic hypertension, then tells you to confirm with a cuff — FDA-cleared as a screening tool, not a BP monitor. It catches under half of hypertension cases, on purpose, to keep false alarms down. Obviously, engineering for lots of false negatives kind of limits the utility.

Signal Ring is the newest entrant, and what got me thinking about the topic. It takes the optical approach, like Samsung and Hilo, but claims to need no cuff calibration at all — more on that below.

Clearance versus accuracy

FDA clearance is not proof of accuracy. Hilo cleared by comparison to an existing device (i.e. by the 510(k) pathway), not by passing an accuracy standard — though it did validate separately, against ISO 81060-2, the standard for cuff devices. That standard checks a device at rest against a reference.

Unfortunately, and importantly, it doesn't test the two things you buy a cuffless wearable for: following your pressure as it rises and falls, and staying accurate over time without a fresh cuff calibration. A recent  review in Hypertension Research (highly recommended) traces the problem to the method itself: most cuffless devices need calibration against a cuff, and their ability to track BP changes is constrained by how stable that calibration is and when it was last done.

Why validate a cuffless device against the cuff standard? Because . . . we don’t yet have a standard for consumer continuous cuffless devices. There is a newer standard for continuous cuffless devices used in the operating room, ISO 81060-3, and validated using an arterial line, but that doesn't fit a device worn in daily life.

When will we have a standard? Hopefully soon. One is being written; but until it’s available, a company can only validate against the cuff standard.

Not an ideal situation, and The American Heart Association's 2025 scientific statement said it plainly: cuffless devices aren't yet accurate enough to diagnose hypertension or guide treatment.

Back to Signal Ring

Signal Ring is the most aggressive entrant on the board. Unlike Samsung and Hilo, Signal doesn’t need calibration — the most annoying part of using these devices — entirely.

Signal is also making the boldest claim in the category, which puts it at the opposite end from Apple. Apple claims almost nothing — screening only, no actual BP number. That’s easy to clear, and a nice addition to their already market-leading product that says "we're working on it". Signal has moved to plant their flag, claiming the first to track BP accurately and continuously, no cuff, no calibration.

Bold claims are nice, but so is data. And so far Signal has published no validation study, met no standard, and holds no FDA clearance.

Bloomberg reports that the company behind the ring — which it identifies as Vital Signals, a name that appears nowhere on the product's own site — has a separate "medical-grade" version in trials at Stanford and three other sites, and claims accurate readings across thousands of participants. But that isn't the device shipping in October, the trial data isn't published, and there's no validation for the consumer ring at all. In a hands-on demo, Bloomberg got some inaccurate readings, which the founder attributed to fit.

Continuous versus periodic: why it matters

The interesting question isn't where blood pressure gets measured. Home cuffs have been around for decades — one of the first elements of the decades-old Monitoring Migration that I've written about. No, the interesting question here is what continuous measurement might show us that periodic BP can't. That's what happened with heart rate, pulse oximetry, and sleep: the value wasn't only in moving them out of the clinic, it was in the constant stream, which revealed patterns a few daily (or annual!) readings missed. Blood pressure varies more than almost any of them — across the day, with stress, with sleep, with medication. I'm betting that continuous BP will show us important things that periodic measurement hasn't.

Maybe we will get that from the Signal Ring, though it's impossible to say at this point because the company hasn't shown us any data. Hopefully that changes before the hardware comes out in October.

In March I wrote that OpenEvidence, now used by roughly two-thirds of US doctors (up from about 40% when I wrote), and its specialized competitors — UpToDate Expert AI, DoxGPT, Glass Health, Vera Health, among others — had never all been independently tested against each other.

Three months later, two head-to-head studies of AI-for-medical-purposes arrived within weeks of each other. They reached opposite conclusions. One includes a comparison of OpenEvidence and UpToDate's AI, but neither answers the question every practicing physician has: which of these should I use?

Two studies, opposite conclusions

Last month, an NYU team led by Krithik Vishwanath and Eric Oermann published a comparison in Nature Medicine: OpenEvidence and UpToDate Expert AI against three general-purpose AI (GPT-5.2, Gemini 3.1 Pro, Claude Opus 4.6), tested on 500 MedQA licensing-exam questions, 500 HealthBench items, and 100 real physician queries, with twelve clinicians doing blinded review against fixed answer sheets. The general-purpose AI won every evaluation. Amazingly, the medical AI performed roughly as well as Google's AI Overview — the free feature above your search results.

Two weeks later, a UCSF-led team headed by Jean Feng found ... the opposite. Using 620 real point-of-care questions from the OpenEvidence platform, 149 practicing physicians across 36 states graded AI responses relevant to their specialties, with data collection run and paid for by OpenEvidence. The tested AIs: OpenEvidence, Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5.

And OpenEvidence both funded and helped to design the Feng study — making it unsurprising that OpenEvidence came out on top by a wide margin.

The table highlights the major flaws in both studies, including: the NYU study was small; the UCSF study was designed and paid for by OpenEvidence.

The question neither study fully answered

The NYU study, for all its limitations, does contain an independent medical-vs-medical comparison. OpenEvidence and UpToDate Expert AI were both included — a genuine head-to-head between two medical AIs, run by a team with no stake in either. Both scored near Google's AI Overview, and pretty close to each other. That is the first independent comparison of two of these products.

But: DoxGPT, Glass Health, and Vera Health were not included, and remain untested against anything. So we now have a comparison of UpToDate vs OpenEvidence that was close to a tie (and close to a tie with generic Google AI), and no evaluation of three other important medical AIs. That's slightly better than when I wrote in March, but still short of what a physician choosing among five would need.

Why the study you want doesn't exist

Short answer: the AI companies don't profit from the answer.

Pharmaceutical companies rarely test their drugs against competitors — head-to-head trials happen mainly when a company is confident it will win. Losing one costs market share, so the comparisons clinicians most need don't get run. The same logic governs clinical AI. OpenEvidence will collaborate on a study where OpenEvidence beats ChatGPT. Wolters Kluwer will fund one where UpToDate shines. Neither will fund OpenEvidence versus UpToDate versus DoxGPT versus Glass Health — unless they can have a strong hand in guiding the design, as in the Feng study.

The same incentive explains why outsiders can't run the comparison themselves: the medical AIs are enterprise products whose terms of service discourage benchmarking and offer no access at the scale a real study needs. A determined team can still test them by hand, as the NYU group did with two of them — but not across the number of tools or the volume of questions the full comparison requires. The general-purpose AI — ChatGPT, Gemini, Claude — have open APIs, so anyone can test them, which is why every study includes them.

Plus:

Expert participation is expensive. The UCSF study needed 149 specialty-matched physicians to generate 1,156 ratings, and the authors still concede that they don't have enough data to support conclusions within any single specialty. Getting enough ratings to answer that would be a million dollar activity, and NIH won't fund it.

The field is moving too fast. Between the NYU preprint and its Nature Medicine publication, the model lineup had to be updated — GPT-5 became GPT-5.2, Claude Sonnet 4.5 became Claude Opus 4.6. A rigorous evaluation takes months or years; by the time it's published, every system tested is obsolete.

What would a useful study look like

So what would a study look like that could actually help doctors choose which AI to use?

1. Neutral questions — not drawn from any competitor's platform — spanning diagnosis, triage, and management, not just point-of-care lookup.

2. All five medical AIs — and any newcomers — alongside the latest versions of the general-purpose AI.

3. Specialist panels large enough to grade each answer as acceptable or not, so questions with several right answers still get scored fairly.

4. A standing setup that reruns as models change, instead of a snapshot obsolete by publication.

The one missing ingredient is a funder with no dog in the fight — a medical society, a large payer, a health system consortium, anyone whose interest is the answer rather than a particular winner. Or an appropriate regulator that, like FDA, would dictate design and audit the data.

Every other part of medicine demands more. A new statin can't claim superiority over an old one without a head-to-head comparison. A new diagnostic test has to report its sensitivity and specificity against a reference standard. A clinical AI shaping millions of decisions a month? That just needs a good marketing team.

And until that changes, no study will answer the question doctors actually need: which AI is best for my practice?

This is close to what Doctronic has been doing in Utah — though Doctronic handles refills and UpDoc adjusts doses, orders the confirming lab, and documents the visit, all without an appointment.

In both, AI is doing what only a licensed prescriber used to., and in both cases it is — for the moment — happening inside the healthcare system. Sustaining innovation for the healthcare system, potentially disruptive for the medical profession.

Of course, if the pilots go well, it’s not hard to imagine future AI systems prescribing medications rather than adjusting or renewing them. And that won’t improve the healthcare system OR the medical profession, it will render both less relevant to our health.

This is what we hear from institutions, too: the AAMC projects a shortfall of up to 86,000 physicians by 2036, HRSA up to 187,000 by 2037.

And from Congress: the Resident Physician Shortage Reduction Act would add 14,000 Medicare-funded residency slots over seven years — spending billions to expand physician training, backed by the AAMC, the AMA, and the hospital associations.

But do packed clinic schedules in coastal cities really mean we don’t have enough doctors?

The Boston problem

In AMN Healthcare's 2025 survey of physician appointment wait times, Boston had the longest average wait in the country: 65 days for a new patient appointment, against an average of 31 days.

Massachusetts also has more physicians per capita than any state in America — 485 per 100,000 against a national average of 304. It leads the states on direct-patient-care physicians, too, at 371 per 100,000 against a national average of 255.

So . . . the place with the most doctors has the longest waits.

AMN somehow offers this as evidence of a doctor shortage. Which makes zero sense: if more doctors meant shorter waits, Boston would be the easiest place in America to make an appointment.

The same pattern holds across all fourteen metro areas: when you plot each city's wait time against its supply of direct-patient-care physicians, it just doesn’t show anything like “more doctors = shorter waits”.

[Note: I left Washington, DC, the fifteenth metro AMN surveys, off the graph. DC is a city with no state of its own, and its physicians serve the whole region, not just DC residents. This inflates its per-capita density beyond any other metro.]

If low supply drove the booking delays, the densest metros would have the shortest waits. Instead supply explains essentially none of the variation. Whatever is behind these waits, a raw shortage of doctors ain't it.

A wait time is a ratio

How long you have to wait for an appointment reflects a ratio: the number of people seeking appointments, divided by the number of appointment slots on offer.

The numerator — people seeking appointments — depends on population, on how sick that population is, and on something the shortage debate ignores entirely: how many health tasks still require an appointment at all. Every task that moves to a chatbot, a wearable, a home test, or a DTC platform removes appointment-seekers from the numerator without anyone in a clinic noticing.

The denominator — available slots — is more complicated than "number of doctors." How many slots exist depends on several things:

• How many clinicians are practicing (physicians, NPs, PAs)

• How many hours each one works, and how many of those hours go to bookable visits rather than documentation, inbox messages, and prior authorizations

• How many points of access exist, which depends on clinic openings, closings, and consolidation

• How long each visit runs, since the same clinician-hours yield fewer slots when each patient takes longer — and as the simple visits shift to consumer tech, the ones left in the clinic are likely to be more complex

A long wait tells you the ratio is high, but doesn’t tell you which term moved. That's the critical mistake in the doctor-shortage consensus: assuming the numerator (demand) is rising.

Busy clinics don't prove increasing demand

Every single time I mention demand migrating out of the healthcare system to a doctor in practice, they tell me "well, my clinic is packed."

But busy clinics don’t prove that demand is high. Busy clinics happen whenever demand outpaces capacity, and that can happen whether demand is rising OR falling — because capacity can rise or fall, too.

Take two clinics with identical month-long waits:

• In the first, visit demand is falling — patients are handling more themselves, asking an app the questions they once booked a visit for — but clinical capacity is falling faster: retirements, burnout exits, reduced hours, closures, consolidation. Fewer doctors have to deal with a still-larger pool of patients — a shrinking restaurant district can still have lines out the door.

• In the second, visit demand is rising with an aging population, and capacity is falling for the same reasons. Demand up, capacity down.

Both clinics are slammed. Both have a long wait. Both waiting rooms look identical. And only the second clinic represents the demand-rising story my busy physician colleagues — along with the press, and the government, and the trade organizations — assume. The busy clinic is evidence that demand currently exceeds capacity — and that’s it. It can’t tell you which one moved, and so it can’t, by itself, support either my claim that demand is leaving OR the common claim that supply is insufficient.

Of course, genuine localized shortages can still exist in certain specialties or underserved areas. I’m talking about the national picture.

What the data show

The denominator: slots

Total physician supply is growing, not shrinking. The US had 1,032,365 active physicians in 2024 — 304 per 100,000 people, up from 278 per 100,000 in 2018. America has more doctors than ever — per capita and in absolute numbers.

Primary care supply is roughly flat per capita. The Milbank scorecard puts primary care physicians at about 67 per 100,000, down slightly over the past decade, with total primary care clinicians (including NPs and PAs) at 103.8 per 100,000 in 2022, also down slightly. A modest decline — not a collapse.

The access infrastructure is consolidating fast. Between 2018 and 2026, hospitals and corporate entities acquired 85,000 physician practices; the number of independently owned practices fell by 81,100, and only 18% of physicians now work in physician-owned practices. Acquisition doesn’t always mean closure — but acquirers typically prune the acquired organizations. In 2024 and 2025, corporate entities bought 8,000 more practices while growing the physician headcount inside them by less than 1% — prompting the Physicians Advocacy Institute, which tracks these acquisitions, to ask whether owners were shedding doctors after buying the practice.

The attempts to add capacity are failing. The retail companies that promised new primary care slots are retreating: Walmart Health shut all 51 of its clinics and Walgreens' VillageMD closed about 160. And of course each closure adds its patients into the remaining clinics' schedules.

But nobody counts the slots. No agency, registry, or dataset tracks the total number of primary care or specialist appointment slots in America over time. We count doctors. We count visits that happened. We do not count the capacity offered, or how much of each physician's time still consists of a bookable slot rather than admin duties. So the denominator itself is a guess.

The numerator: appointment-seekers

Demand for traditional visits is falling. Ganguli et al. found primary care visits per capita fell 24% from 2008 to 2016, with roughly three quarters of the decline unexplained by shifts to other clinical settings. The authors speculated that part of the decline could be because of Dr. Google: patients using the internet to answer medical questions.

And that was before ChatGPT, before continuous glucose monitors at CVS, before 14 million US adults skipped a provider visit in a single month after getting advice from AI — a finding I covered in April.

In any case, if demand were genuinely surging against fixed supply, per-capita visit rates would be rising. But they're falling (at least in primary care).

And nobody counts where the demand went. We have no measure of how much health need is now met without a provider at all — answered by an AI chatbot, tracked by a wearable, resolved by a home test. A person who manages their condition through an app and never books a visit is invisible to every available dataset. So we cannot say whether falling visits mean demand is falling — vs migrating out of the clinic.

The models assume the answer

The shortage projections can't resolve this, because they assume their own conclusion. Both the AAMC and HRSA models estimate future demand the same way: they take current per-capita utilization rates and multiply by demographic projections.

Neither model contains a variable for demand leaving the clinical setting. Neither model accounts for the Knowledge Migration already visible in Ganguli's missing visits, or for Monitoring Migration to the wrist, or for Diagnosis Migration to the chatbot. Same thing for Demand Elimination via GLP-1 medications shrinking the need for everything from bariatric surgery to diabetes visits to knee replacements: the models assume these things don’t exist.

A model that defines demand as "what people did in 2021, forever" will project a shortage and rising demand in any future where the population is getting older — including a future where half the appointments have moved to channels the model can't see.

A recent JAMA paper shows the same assumptions. Four physicians propose financing primary care as a public utility, citing as evidence of failure that more than a third of US adults lack "a usual source of care." But “usual source of care” is a measure of utilization, not necessarily a measure of demand. Those without a usual source of care include both people who need care and can't get it — and also people who are getting what they need through an app or a wearable or a Google search.

Once again the healthcare system seems incapable of imagining, or measuring, the possibility of demand reduction, or migration of health activities out of the healthcare system.

This is the same blind spot I described in the $50B measurement piece: the system measures the healthcare delivered inside its own walls and mistakes that for the whole of how people stay healthy — missing the self-care increasingly happening outside it, where no provider is involved at all.

Diagnosis before cure

If you look at these findings together — per-capita physician supply rising, primary care supply roughly flat per cap, per-cap visits dropping, clinics consolidating and slots unmeasured — the pattern doesn’t allow us to distinguish between the demand-falling scenario and the demand-rising scenario.

Which brings us back to Boston. The city with the most doctors and the longest waits is exactly what you'd expect in a system where academic medical centers have absorbed the practices, the absorbed practices offer fewer slots, the simple visits have migrated out, and the complex ones that remain take 40 minutes each.

Does America have a doctor shortage? The data we have show more doctors than ever serving fewer traditional visits than ever through fewer access points than ever. And as I argued last month, the set of tasks that requires a physician is shrinking while we debate how many physicians to train.

So authorities are proposing that we spend billions training more doctors for a system whose patients are leaving — but we haven't yet bothered to quantify either the currently available clinician slots or the extent to which demand is departing.

I suggest we consider spending a tiny fraction of those dollars measuring both before we commit to a cure that doesn't match the illness.

AMN Healthcare's 2025 survey tracks wait times across six specialties in fifteen metros. I plotted each metro's average wait time to get an appointment against its physician supply (using AAMC statewide physician supply as a proxy for metro physician supply) and found . . . no correlation.

In fact, Massachusetts has the highest physician-to-population ratio of any state but Boston has the longest wait of any metro surveyed — 65 days against a 31-day average. AMN notes the paradox in places like Boston, but still concludes the core problem is simply not having enough doctors:

“if patients are having difficulty scheduling appointments in these highly populated areas, it can be assumed that access to physicians may be even more problematic in areas with fewer physicians.”

I guess if the only tool you have is a stethoscope…

Of course, this graph doesn't prove no shortage exists, but it demonstrates that it’s a lot more complicated than simply “more doctors is the solution to long wait times.”

More on this on Monday.

Starting June 16, Oura is rolling out Counsel Health inside the Oura app — members in 43 states can ask a health question, get advice from a medical AI, and pull a licensed physician into the conversation when needed. The device stops being a tracker. It becomes the front door to care.

I've been arguing for two years that healthcare tasks are migrating out of the exam room — Knowledge, Diagnosis, Treatment, Monitoring, Interpretation. A single $399 consumer product is now assembling all five at once. The ring monitors. The AI interprets and diagnoses. It answers the questions people used to save for an appointment. When a prescription or referral is needed, Counsel routes you to a physician. The physician is still there — but he’s now a feature you summon immediately rather than a visit you schedule two months in advance.

This is not one company's bet

Whoop announced on May 8 that it will roll out on-demand video consultations with licensed clinicians this summer, a paid add-on. Each clinician opens the consultation already knowing the member's recovery scores, resting heart rate, and months of strain data.

Whoop and Oura both position these clinical features as supplements to primary care rather than replacements. That wording is deliberate. In July 2025, the FDA sent Whoop a warning letter over its Blood Pressure Insights feature, saying the company was marketing an unauthorized medical device intended to diagnose, treat, or prevent disease. Six months later the FDA issued new guidance permitting optical blood pressure sensing in wellness devices, provided they make no medical-grade diagnostic claims. The "supplement" framing is what keeps these products in that wellness lane and out of the medical device pathway — the lane the industry helped define.

Oura has been here before. In June 2025 it partnered with Maven Clinic to route ring data into providers' treatment plans. The Counsel deal a year later is the more aggressive move — it brings the provider into Oura's app rather than sending Oura's data into the provider's workflow.

The pattern is roughly the same each time: a continuous biometric stream, an AI layer that interprets it, and a clinician available on demand. That sounds a lot like a health system, doesn’t it: it measures you, it interprets the measurement, and it connects you to a doctor. The wearable companies have reassembled those functions outside the hospital, on a subscription, for the price of a copay. Oura and Whoop might be leading this, but others will follow and it is too soon to pick winners.

What this means for health systems

For a hospital CEO, the temptation is to file this under consumer wellness and move on. The rings aren't FDA-cleared diagnostic devices, the consults are explicitly positioned as supplements, and the volume is still small. This is beside the point.

The front door is a sales funnel: it directs patients to higher-level care. That higher-level care is where a health system makes its money. The office visit rarely paid much; the referrals it generated did — the cardiology workup, the MRI, the surgery. A ring that answers "is this chest pain worth worrying about" now sits at the mouth of that funnel, deciding which questions become referrals and which never leave the app.

Health systems understood the value of the front door years ago. The whole digital-front-door movement — the patient portals, the symptom checkers, the virtual-first plans, the apps that promised to be the first place a patient looked — was an attempt to own first contact before someone else did. The instinct was correct. The execution lost to companies that started with better data, better software design, and no waiting room. A patient opens the hospital portal to request an appointment three weeks out. She opens Oura and has an answer in seconds, with her own biometrics already in the chat.

The wearable companies aren't trying to build hospitals, but they are working to capture a lot of what we traditionally think of as healthcare. For someone who is basically healthy, a system like this may eventually handle the majority of their healthcare interactions: answering a question about a medication, flagging a trend worth watching, routing to a physician when the AI hands off. These are largely interactions that pass information back and forth.

The clinical system will have trouble seeing this shift. Most of these interactions generate no bill — the monitoring, the AI interpretation, the health coaching have no CPT codes. The in-app telemedicine visits are direct-pay subscription add-ons, following the same cash-pay model as Ro and Hims — deliberately outside the insurance system. The dominant DTC telehealth players have always worked this way: margins approaching software, no reimbursement negotiations, no claims data. None of this appears in the databases that health system planners and regulators depend on. The demand leaves the system before it can be counted. Health systems won't miss it until they feel it in their census numbers — and by then, it will likely be too late.

In Silicon Valley: We believe AI will be practicing medicine momentarily. Also in Silicon Valley: The technology used to notify the on-call trauma surgeon that an injured patient is arriving. Yes, I concede there is a gap in expectations vs reality.

Obviously, she's right that we live in a world of amazing advances in medical AI, while at the same time much of healthcare is mired in old tech. Point conceded.

But "will be practicing medicine momentarily" is wrong in using the future tense. AI is practicing medicine now. It has been, legally, since at least 2018, when the FDA authorized an autonomous system to diagnose diabetic retinopathy with no physician interpreting the result. And AI now renews prescriptions in Utah under state authorization, again with no physician signing off. In both cases the company producing the AI, not a doctor, carries the malpractice policy.

When I pointed this out, Dr. Palm answered:

I think we can debate the semantics of whether these tools are really "practicing medicine" yet. Let's not sell ourselves short by setting that bar so low.

But when I diagnose something in clinic, or prescribe a medication, or refill one, I am definitely practicing medicine — and if you did any of that without a medical license you'd be arrested.

So if a machine does the exact same things, how is that semantics?

What the practice of medicine actually means

Every state has a medical practice act, and they all include the same short list of acts a license exists to gate: diagnosing disease, prescribing or administering treatment, and performing procedures, generally for compensation. California's is representative — Business and Professions Code §2052 makes it a crime to diagnose, treat, or prescribe for any condition without a license. The Federation of State Medical Boards' model language reads the same way. None of these statutes mention empathy, a white coat, or bedside manner.

That's the bar — not mine, not Silicon Valley's. And three things have already crossed it.

The line has been crossed

LumineticsCore, formerly IDx-DR, received FDA de novo authorization in April 2018 — the first clearance ever granted to an autonomous AI diagnostic system. A technician captures retinal images; the software returns a diagnosis of more-than-mild diabetic retinopathy "without the need for a clinician to also interpret the image or results," in the company's own words. The act that once required an ophthalmologist's judgment now requires none. Medicare reimburses it. Digital Diagnostics carries malpractice insurance on the output and assumes liability for the diagnosis — the company stands behind the result instead of an ophthalmologist.

In January 2026, Utah's Office of Artificial Intelligence Policy authorized the first state-approved program letting an AI system write prescription renewals. A patient verifies identity, confirms a prescription history, answers clinical questions, and — for medication classes past the program's validation threshold — Doctronic's AI issues the renewal and sends it to the pharmacy, with no physician involved. The pilot is limited: 191 routine medications, no controlled substances, physicians reviewing the early cases in each drug class. But once a class clears that review, the AI renews prescriptions in it without physician input. Doctronic holds a malpractice policy binding the system to a physician's standard of accountability, because the system does what a physician would normally do.

Aidoc and Viz.ai are the contestable case. Their FDA-cleared software scans medical images for signs of a stroke or a blood clot in the lungs, and when it finds one, moves that patient to the front of the radiologist's list. A radiologist still confirms the finding, so the machine isn't making the final call — but it is reading the image and deciding who gets seen first, on its own judgment. That is a piece of triage, work that has always belonged to clinicians. I'd call it a partial crossing and let the reader argue — but the first two rows are crystal clear.

The retreats

In my experience, once the examples are on the table, the objectors keep moving the goal posts, redefining "practice" to exclude whatever the machine just did.

It's only narrow. So is a dermatologist. Specialization has never disqualified an act from being the practice of medicine.

A physician is still ultimately responsible. Not with LumineticsCore or the Doctronic pilot in Utah.

It isn't making complex judgments. If a diagnosis that previously required a physician with at least 8 years of post-college education wasn't complex, why did it require that specialist? And if a machine starts making the same diagnosis, did it suddenly become less complex?

It's just a tool. Tools don't carry their own malpractice insurance. A company can only get medical malpractice insurance — as opposed to ordinary product liability — if that company's product is legally practicing medicine.

Why so much resistance

The resistance is understandable. Medicine's professional identity rests on the premise that certain acts require a licensed human.

Up to about 3 years ago, this identity manifested to me in the form of visibly angry physicians during the Q&A after my keynotes, insisting that "you can never replace the judgment of a physician with a machine."

I don't get as many of those questions anymore, as physicians seem more and more resigned to the power of AI in performing medical tasks.

But physician groups are still resisting now, under the guise of patient safety. For example, when Utah's pilot went live, the state medical licensing board demanded its suspension, writing that "there is a reason prescription refills require physician authorization," and in the process conceding that machines can write prescriptions — though, in their view, not safely.

The 2018 FDA authorization of LumineticsCore (then IDx-DR) drew no such letter: it does one narrow thing in one specialty, and few outside ophthalmology had cause to protest when a small, well-paid field lost a routine screening task. Prescription refills reach all of primary care, and the alarm has grown accordingly.

And it's just getting started

The acts that define medical practice are migrating to machines — under FDA authorization, with vendor liability, billed to Medicare. It began with two of the most consequential acts a physician performs, and it will not stop there. And this is only the migration happening inside the clinic — the larger one is just starting, as more and more physician tasks move not to the AI a hospital buys but to the AI a patient already carries. That's a subject for another piece. Healthcare is getting better, and getting smaller, and the smaller part begins with the acts physicians used to own.

Apple's Health app genuinely prices interpretation at zero — no subscription required, ever. The Fitbit Air's AI interpretation layer, Google Health Coach, requires Google Health Premium at $9.99 a month. After the included three-month trial, the five-year cost of Fitbit Air plus AI interpretation is approximately $668 — not $99.

[Note: if you pay for Gemini, you get the health AI included. Trying to find out if that’s the $7.99 a month Gemini tier, which would be slightly cheaper than the $9.99 per month, but still way more than Apple.]

The corrected chart is below, with Apple a full 40% less for five years than the new "inexpensive" Fitbit Air.

Apple's model remains the outlier: a one-time hardware purchase that includes AI interpretation of your biometric data at no ongoing cost. Every other player in this space — Google, Oura, Whoop — charges a subscription for the interpretation layer. The monitoring is cheap; the meaning costs extra.

That distinction is exactly what the Interpretation Migration is about. Apple eliminated the recurring cost of interpretation. Google didn't.

The FDA just approved Epiduo (adapalene 0.1% / benzoyl peroxide 2.5%) for over-the-counter sale to anyone 12 and up. Until last week, this combination was prescription-only — meaning an acne patient needed an appointment, a diagnosis, and a script to get it. This summer it shows up next to the face wash at Walmart, Target, CVS, and Amazon.

The FDA didn’t approve this because the drug got safer — it approved it because 15-plus years of use proved the prescription gate was no longer needed. Adapalene by itself made the same switch in 2016. Now the fixed-dose combination follows.

So a therapy that required a clinician to authorize it becomes a therapy the patient selects (maybe with the help of AI), buys, and manages alone. No visit. No diagnosis. No prescriber. The dermatology visit for routine acne — one of the most common reasons adolescents see a doctor — has one less reason to happen.

The Treatment Migration continues.

Two findings stand out from the report. When the AI recommended a renewal, the reviewing physician agreed 91% of the time on first review, rising to 97% after a second physician weighed in on the 9% of disagreements — a level of agreement that sits inside the published range of human-to-human disagreement for the same task, against a reference human error rate of 5–12%. On the other side of the ledger, the AI escalated 28% of cases to a human rather than recommending a renewal, and in 31% of those escalations the reviewing physician judged the AI had been overly cautious.

The AI's recommendations are statistically indistinguishable from a second physician's. And when the AI defers to a human, it defers more often than a human would.

The standard objection to AI prescribing is patient safety. Five months of Utah data show the safety risk in this pilot is not AI overreach. It is AI underreach — appropriate for a phase-one system, and welcome from a regulator standing up the first program of its kind, but a limitation on access if it persists.

This matters for Treatment Migration. Treatment Migration does not require AI to outperform physicians. It requires AI to perform within the band of acceptable human variability at a cost and convenience humans cannot match. Utah just published the first real-world data showing that bar has been cleared. The pilot covers 192 medications at $4 per renewal. The reviewing physicians are agreeing with the AI at rates indistinguishable from how they would agree with each other.

The 9% disagreement rate on AI recommendations will get everyone’s attention, but the 31% overcaution rate on AI escalations should get some, too.

Source: Utah Office of AI Policy, May 19, 2026

She is also answering a strawman.

The real question — the one hospital systems, medical schools, and workforce planners need to answer — is not whether AI will “replace doctors”. It is how many physicians we will need as the set of tasks that only physicians can perform are dramatically reduced. Barring some additional development, the answer is many fewer.

The framing trap

"Will AI replace doctors?" is the question physicians keep posing because it sounds like a question physicians can win. Ofri thinks she wins it. Ofri notes that a chatbot doesn't have a 25-year relationship with the patient. And a chatbot cannot read the pattern of a patient's respiration across a waiting room.

Except . . . she is wrong.

First, because AI systems will indeed have longitudinal relationships with all or most of us. An AI connected to a patient's wearables that have been monitoring vital signs, sleep and motion patterns, VO2 max, and more for months, years, or longer knows long before the patient gets to the waiting room that there is a problem. And has more useful data at its disposal than any physician could.

A small example. My 85-year-old mother lives in another city. Last month her Apple Watch — which along with her iPhone has been monitoring her movement for nearly a decade — alerted me that her daily steps and calories burned had been trending downward for several weeks. A quiet drift in her own baseline that no one in a clinic would have caught, because she hadn't made any appointment. I called her and found that she had hurt her ankle, hadn't mentioned it, and was starting to get back to her usual routine.

No visit needed. No doctor. No billing code.

Ofri's other argument is that AI cannot see what a physician sees across a waiting room. What she fails to recognize is that AI now sees what no physician ever sees: the slow daily drift of an octogenarian's activity, from hundreds of miles away.

Not to mention that LLMs can now literally see whatever you point the camera at, including how the patient is breathing.

Ofri's examples are not a refutation of AI capabilities in health. They are a description of them.

The first to go

Meanwhile, the bread-and-butter healthcare encounters are not the complex "Anna Karenina" case she describes. They are quick, transactional — and increasingly happening outside the clinic, the first to migrate.

The actual disruption underway does not require AI to "replace doctors", it just requires AI and other technologies to provide a solution outside the clinic for the many common reasons a patient shows up. Reassurance that a cough is just a cold, informed by a wearable showing normal heart rate and oxygen saturation. A photo of a rash or mole triaged by an AI visual model. Routine titration of a blood pressure medication or statin guided by a connected cuff and an AI that has the patient's daily readings.

Each time such a solution arises, an encounter — and the revenue attached to it — is pulled away from the doctor.

And as AI grows more capable and the data captured by our personal technologies becomes more comprehensible, better understood, and more effectively scaled throughout society, it will pull more and more sophisticated activities out.

It's already happening

This is not simply a theory. Gallup and West Health found that roughly 14 million U.S. adults skipped a provider visit after consulting an AI tool — a finding I unpacked here, pointing out that translates to 168 million missed appointments per year, completely invisible in claims data. Rock Health's 2025 Consumer Adoption Survey put consumer AI chatbot adoption at 32%, with 18% of users self-adjusting medications based on what the bot told them. And Ganguli and colleagues had already documented a 24% decline in primary care visits between 2008 and 2016 — three-quarters of which was invisible in standard claims data, and which they felt was at least partially due to patients using the internet to answer health questions.

This is the Monitoring Migration, the Knowledge Migration, the Diagnosis Migration — happening now, at scale, with or without anyone's permission. ChatGPT health answers the question. A wearable flags the arrhythmia. A DTC platform writes the GLP-1 prescription.

Ofri's essay does not engage with any of this. It can't, because the framing she has chosen assumes the patient is in the exam room to begin with. But a rapidly growing number of them are not.

Even the long relationship is at risk

The 25-year-relationship version of Ofri's argument might be slower to migrate: the patient she has known across two decades of clinic visits, whose gait and heart murmur and family crises she carries in her head.

But Ofri's assumption that long relationships and contextual wisdom are uniquely human territory just doesn't hold. The consumer tools assembling around this are not yet a single product, but the pieces are all shipping now: wearables capturing continuous physiology (Apple Watch, Whoop, Oura, Dexcom's OTC CGM), symptom and mood logs in Apple Health and dozens of apps, lab-trend services like Function Health, and conversational AI that now retains memory across sessions. The integration layer — one agent that synthesizes all of it into clinical reasoning about this specific person — is the obvious next product.

A patient running that near-future set of tools for a year is in a relationship, too. Their IT stack has continuous data, total recall, and no 15-minute visit cap. It knows when their resting heart rate drifted, how they slept the week of the family crisis, what their A1c trend looks like, and what they said about their estranged child eighteen months ago. Ofri's 25-year relationship is built on episodic snapshots a few times a year — quarterly at best, annual for most patients. The integrated consumer record will have minute-by-minute data over years. My mother's decade of Apple Watch data, captured without her ever thinking of it as a medical record, is already a longer continuous physiological dataset than any clinician's chart on her.

AI can't do X — except it can

It is worth noting how often the "AI cannot do X" argument has been wrong. Eric Topol's 2019 book Deep Medicineargued that AI would handle the data and free physicians for the human work AI could never do. For example, AI "cannot empathize with a patient." Then in 2023, Ayers and colleagues published a JAMA Internal Medicine study comparing ChatGPT and physician responses to real patient questions on Reddit's r/AskDocs. Blinded licensed clinicians preferred the chatbot's responses 79% of the time, rating them 3.6 times higher for quality and 9.8 times higher for empathy.

A 2025 replication in the Journal of General Internal Medicine found the same effect — and that chatbot responses scored higher on empathy even when evaluators were told the response came from a physician. By 2024, Topol's framing had inverted: at the Precision Med TRI-CON keynote, he described AI tools that coach physicians on empathy by flagging when they interrupted a patient after nine seconds or failed to show sympathy. To his credit, he updated his position publicly when the evidence shifted.

Getting back to Ofri, there is no obvious reason character understanding, longitudinal context, and clinical wisdom will be exempt from the same technological trajectory.

What's left for doctors

Ofri closes with a call to elevate the medical humanities — to teach the next generation wisdom rather than smarts, because "they can look up whatever they need to know." She is right that wisdom matters more than knowledge. She is most likely wrong that wisdom is the last stronghold for doctors because there is no reason to believe AI cannot gain wisdom (or the convincing appearance of it, as with empathy).

If AI eventually has the empathy, the knowledge, the wisdom, and the continuous data — then the work that most durably remains for physicians is not cognitive or relational. It is physical. Stabilizing a patient after a motor vehicle accident. Suturing a laceration. Setting a fracture. Delivering a baby. Administering anesthesia. Performing surgery. The work that requires hands on a body, in a room, in real time, with legal authority and procedural skill. The work that cannot be done from a phone, a wearable, or a chatbot, regardless of how sophisticated those tools become.

That is a much smaller share of what physicians currently do than the profession wants to acknowledge. It is also more generally defensible — and not because AI has reached some intrinsic limit on judgment or empathy. The physical category holds because dissolving it requires as-yet-unrealized advances not just in AI but in robotics.

How many physicians will we need?

The strawman is whether AI will replace doctors. The real question is how many physicians the system will need once the tasks only physicians can perform are dramatically reduced.

The cognitive and relational tasks are migrating to consumer AI, wearables, and DTC platforms. The empathy gap closed in under five years. The longitudinal record is being assembled outside the clinic. What remains is the physical work — stabilization, procedures, surgery — and a smaller tier of relational practice paid for directly by patients who want presence rather than throughput. Both categories are real. Neither is large enough to sustain the current physician workforce, the current hospital-system economics, the current training pipeline, or the current insurer business model.

So while that is the question Ofri's essay does not ask, it's the one our profession will not be able to avoid much longer.

This isn't a one-off paper from Apple. Their “Apple Hearing Study” launched in November 2019 with the University of Michigan School of Public Health and the WHO. It now has 160,000+ consented participants, is designed to run at least ten years, and has never used a sound booth. Every six months, participants take pure-tone audiometry on their own iPhones, using EarPods or AirPods. The methodology was validated against clinical sound-booth audiometry in a separate study  — and that validation is what allowed the same in-ear audiometry to become an FDA-cleared OTC Hearing Test in 2024, now in 150+ countries.

In effect, Apple used the sound booth to train the sound booth’s replacement.

A second finding, easy to miss: among ~85,000 participants who tested within the WHO "normal" hearing range, 16% still rated their own hearing as fair or poor, and many reported real-world listening difficulties. So it seems that the clinical definition of “normal hearing” may need to be adjusted based on consumer health data (this won’t be the first time this happens, I guarantee).

Two main points matter here from the consumer health tech perspective, and neither is the hearing-mobility link itself (which prior research already suggested).

First, the measurement. Walking speed is one of the strongest predictors of all-cause mortality in older adults — but like VO2, primary care does not measure it. Apple's iPhone has been measuring it passively, for every user who carries the phone, since iPhone 8 — automatically populating the Health app whenever someone walks with the device in a pocket or a waist pouch. The 57,183-person analytic sample is a tiny slice of the deployed sensor base, which runs to hundreds of millions of devices in the US alone. The Monitoring Migration is not theoretical — it is shipping at a scale primary care cannot approach.

Second, the infrastructure. 160,000 participants — no clinic visits. Biannual clinical-grade audiometry — at home. Continuous passive gait, sleep, and heart-rate data — at home. Near-zero marginal cost per enrollee. No sound booths, no field staff phoning participants to schedule them. Michigan epidemiologists designed the protocol and did the analysis; everything in between was the iPhone.

The Framingham Heart Study took roughly 75 years to enroll about 15,000 people across three generations. Apple's hearing cohort is ten times larger and was assembled in five.

The Five Migrations describe healthcare tasks moving from the clinic to the consumer. This study is a reminder that medical evidence production (i.e. research) is moving the same direction — and the clinical system will increasingly find itself using research it did not produce, about activity and indicators it cannot measure, generated by devices it does not control.

More on this soon.

Google launched the Fitbit Air on Wednesday at $99.99 with no required subscription. The Fitbit app is being rebranded Google Health on May 19, with an optional $9.99-a-month AI coach tier on top. Pre-orders opened the same day.

The launch is being covered as Google entering the Whoop and Oura category. I think the more useful framing is that Google joined Apple, years late, in pricing the interpretation layer at zero.

Since 2015, Apple Health and the Apple Watch have given users data plus interpretation. They’re now including VO2 max, heart rate variability, sleep stages, AFib detection, walking steadiness, cycle tracking, fall detection, trend analysis — all included with the hardware. Apple does not charge a recurring fee to look at your own data or to receive basic interpretation of it.

Whoop and Oura are in a different camp. Whoop sells no hardware at all — the strap is a sunk cost the company recovers through a mandatory $199-to-$359-a-year membership. Oura sells the ring at $349 and up, but the free tier shows three numbers, which means the $69.99-a-year membership is pretty much required.

For comparison, an annual physical with basic labs runs roughly $300 cash-pay at a primary care office.

With Google’s latest move, the math looks a bit different than it was just a week ago.

Five years of continuous biometric tracking with interpretation now runs $99 on the Fitbit Air, $399 on Apple Watch, $699 on Oura, $995 on Whoop One, and $1,795 on Whoop MG. The Fitbit Air at $99 — for five years — is cheaper than six months of Whoop at the cheapest tier — and cheaper than a single year’s annual checkup.

To be fair to the annual checkup, that typically includes things wearables don't — physical exam, vaccinations, lab draws and lab interpretation, prescription authority. But on labs specifically, DTC services like Function Health ($499/yr for 100+ biomarkers), Marek, and Hone Health now offer broader panels than the standard physical orders, with results delivered to a phone. And the checkup function — vitals, "let's see how you're doing," general guidance, "your numbers look fine" — is exactly what the platforms now do continuously, every day instead of once a year.

Two takeaways.

The first is for Whoop and Oura. Their total five-year cost runs seven to eighteen times what the Fitbit Air costs. They now have to explain what the buyer gets for the difference, in a category where one of the largest companies on earth is now giving it away. Some of that explanation might be real (Whoop pushes athlete-tier coaching, for example) but most of it is going to be hard to justify.

The second takeaway is for health systems: the annual physical is dying. Ganguli at Harvard told us that back in 2019 and that was before the Interpretation Migration enabled by AI was even hitting its stride. Now that interpretation is moving to scale, patients are dropping out in droves. A lot of the whole reason for the checkup — vitals, "let's see how you're doing," general guidance — is now available continuously, in the background, for $99. And consumers can now track very useful biometrics that were never available at the checkup — VO2 max being the cleanest example.

It’s called VO2 max.

You’ve probably never heard your doctor mention it.

The number

VO2 max measures the maximum amount of oxygen your body can use during intense exercise. It’s a single number — reported in milliliters of oxygen per kilogram of body weight per minute — that captures how well your heart, lungs, blood vessels, and muscles work together. Researchers call the broader concept “cardiorespiratory fitness,” or CRF. VO2 max is how you measure it.

And it may be the single most powerful predictor of whether you live or die that medicine has ever identified.

In 2022, a study of more than 750,000 U.S. veterans found that even modest improvements in VO2 max were associated with a 13 to 15 percent reduction in mortality risk. Regardless of age. Regardless of BMI. Regardless of sex or existing conditions.

A 46-year follow-up of over 5,000 men in Copenhagen found that each 1 ml/kg/min increase in VO2 max was associated with 45 additional days of life.

A 2018 Cleveland Clinic analysis of 122,007 adults found that those with the lowest VO2 max had four times the mortality risk of those with the highest. And the benefit curve never flattened — fitness kept paying dividends well past “fit” into “elite.”

To put that in perspective: the mortality difference between the least fit and the most fit people in that study was larger than the difference between non-smokers and smokers. Larger than the difference between people with and without diabetes, or the difference between people with and without end-stage renal disease.

The recommendation no one followed

In 2016, the American Heart Association published a landmark scientific statement arguing that cardiorespiratory fitness should be treated as a clinical vital sign — measured routinely, right alongside blood pressure, heart rate, and temperature. Their recommendation was explicit: at a minimum, all adults should have their cardiorespiratory fitness estimated every year.

That was nearly ten years ago. Has your doctor ever measured it? Have they ever mentioned it?

Mine never has. And I’m a doctor.

The reason is straightforward. Measuring VO2 max accurately requires a graded exercise test — a treadmill or stationary bike, a face mask connected to a metabolic analyzer, a trained technician, and a patient willing to exercise to exhaustion. The test takes 10 to 15 minutes and costs $150 to $250. Not covered by insurance.

In other words, the single most powerful predictor of mortality we have doesn’t fit inside a 15-minute office visit. So it doesn’t get done.

A 2024 update to the AHA’s original statement acknowledged the problem, noting that despite overwhelming evidence, “inertia exists” with respect to integrating CRF assessment into routine care.

But it isn’t inertia that prevents VO2 max from being integrated into standard checkups. It’s financial reality dictating that there is absolutely no way to turn 15-minute visits into 30-minute visits without sinking the ship.

The migration

While the clinical system was busy not measuring VO2 max, someone else was — on millions of people’s wrists.

Apple Watch has been estimating VO2 max since 2020. It does this passively — using heart rate sensors and GPS data during outdoor walks, runs, and hikes. No mask. No technician. No appointment. Since late 2020, it has also been sending push notifications to users whose cardio fitness is consistently low enough to suggest health risk.

Apple validated its algorithm against gold-standard lab testing with 755 participants. The result: accuracy within about 1.2 to 1.4 ml/kg/min—roughly 4 percent of true value. About one in five Americans now wears a smartwatch or fitness tracker. Garmin, Whoop, Fitbit, and others estimate VO2 max too.

Is a wrist estimate as good as a lab test? No. A 2025 validation study found that Apple Watch underestimated VO2 max by a meaningful margin. Is it better than a gold-standard lab test that never gets done? Yes, yes it is.

But the wearables I’ve mentioned don’t just measure your VO2 max annually, as suggested by AHA. They measure it every time you take a brisk walk. And a wrist estimate taken hundreds of times over months and years is a different animal completely. It captures trends. Trajectory. Response to training. Response to illness. Seasonal variation. For a metric whose clinical value is largely about change over time, a less precise measurement taken continuously is almost certainly even more useful than a more precise measurement taken once a year.

Last month, Apple ran a global ad campaign — Listen to your body. Not everybody — built entirely around the idea that your wrist now knows things about your health that nobody else does, not even your doctor. In essence, the many consumer electronics companies offering VO2 max estimation are running the public health practice that AHA offered but healthcare couldn’t make happen.

Beyond the system

I often talk about the “Five Migrations” of clinical tasks from the exam room to the living room: Knowledge, Diagnosis, Treatment, Monitoring, and Interpretation. But this isn’t a migration, because measurement of VO2 max was never in the exam room. This is not a story about consumer technology doing that isn’t quite as good as something healthcare does, it’s a story of consumer tech doing something healthcare is simply not able to do. Something very important for our health.

Worse, the system can’t even see that it’s happening. Wearable VO2 max estimates generate no insurance claim. No CPT code. No encounter. No entry in the medical record. When a 55-year-old woman gets a low cardio fitness alert on her Apple Watch, starts a HIIT program, and improves her VO2 max by 5 ml/kg/min — reducing her all-cause mortality risk by somewhere around 50 percent — that entire health intervention shows up nowhere in the healthcare system’s data. Her doctor doesn’t know. Her insurer doesn’t know. CMS doesn’t know.

For those healthcare execs and health policymakers that only count what the system measures — which is almost everyone — it didn’t happen.

This is the same blind spot I’ve written about before. Per capita healthcare spending growth has fallen from 83 percent per decade in the 1960s to 28 percent today. Primary care visits declined 24 percent between 2008 and 2016, with three-quarters of that drop invisible in claims data. The system is getting smaller. But the measurement apparatus can only see what happens inside the system. And increasingly, the most important things are happening outside it.

What this means

I check my own VO2 max on my watch. I’ve never had it measured by a physician — and I’ve never measured it for my patients. That’s an awkward admission, but it’s an honest one. And I suspect it’s true of most doctors and most patients.

The AHA was right in 2016. VO2 max might be the most important vital sign we have. They were right to recommend annual measurement. But they were recommending that a system built around 15-minute visits, fee-for-service billing, and passive diagnostic tools should suddenly start administering exercise tests to every adult in America. That was never going to happen.

What happened instead is that a consumer electronics company built the measurement into a device that 100 million people already wear, and started nudging the least fit among them toward behavior change. Nobody planned this as a public health intervention. No committee approved it. No pilot study preceded it.

And VO2 max isn’t the only example. It may not even be the best one. Clinical medicine was built around episodic encounters — because for most of its history, that was the only access it had to patients. Consumer technology doesn’t have that constraint. And in case after case, it’s starting to do things the clinical system can’t.

More on that soon.

Last week I wrote about an 18% figure from Rock Health — the share of AI chatbot users who say they have adjusted their own prescription medications based on AI advice. The day before that, I wrote about Rock Health's finding that 32% of U.S. adults now use AI chatbots for health information. ALSO last week, I flagged a new Gallup/West Health projection: 14 million U.S. adults skipped a provider visit in the past 30 days after receiving AI-generated health advice.

Two separate surveys, two separate research teams, two separate samples. One phenomenon.

Patients are doing things outside the healthcare system that the healthcare system used to do for them. They are looking up symptoms, adjusting medications, interpreting results, and deciding whether to come in at all. And the system they are routing around has no way to see any of it.

The Gallup data is the most important of the three for anyone doing strategic planning. Rock Health told us patients are using AI. Gallup tells us patients are substituting it for visits, at scale, right now.

The numbers

Gallup surveyed 5,500 U.S. adults between late October and late December 2025. 25% said they had used AI for health information or advice. Among recent users, 14% said the AI advice led them to skip a provider visit in the past 30 days. Gallup projected this to roughly 14 million U.S. adults.

The other findings:

• 59% of recent AI users research before visiting a doctor

• 56% research after visiting a doctor

• 14% used AI because they couldn't pay for a visit

• 16% because they couldn't access a provider

• 21% because they felt dismissed by a provider in the past

• Only 4% strongly trust the accuracy of AI health information

That last number is the most important.

The trust floor

Patients are not skipping visits because they believe AI is better than their physician. 96% of users do not strongly trust AI health information. They are using it anyway, and some are making care decisions based on it anyway, because they believe the alternative is worse (!).

This is the demand-side mirror of a pattern I've written about from the supply side. Ganguli, Lee, and Mehrotra documented a 6–25% decline in primary care visits from 2008 to 2016 across five independent data sources. In their 2019 JGIM paper, they proposed three mechanisms: reduced patient ability or desire to seek care, shifts in how primary care itself is delivered (teams, non-face-to-face modalities), and patients going elsewhere. That "elsewhere" included specialists, urgent care, retail clinics, and telemedicine — and, explicitly, self-care:

"Growing access to Internet-based health information, symptom checkers, and virtual patient communities may preclude seeking formal medical care altogether."

That was 2019. Pre-ChatGPT. Pre-GLP-1 consumer access. Pre-most of what we now call consumer health AI. Ganguli and colleagues had already identified online self-directed care as a driver of the visit decline — a formal entry in their mechanism table, under "Self-care."

What the Gallup data now captures is that same behavior, accelerated by tools vastly more capable than the 2013-era symptom checkers Ganguli mentioned. The 14 million skipped visits aren't a new phenomenon. They're the continuation of a trend Ganguli had already identified seven years ago, now running on dramatically better infrastructure.

What claims data misses

A visit that never happens generates no CPT code, no RVU, no line item. Nothing enters the revenue cycle. Nothing enters utilization reports. Nothing enters the demand forecasts that health systems, payers, and workforce planners use to project capacity three and five years out.

The AAMC's 86,000-physician shortage projection assumes current utilization patterns hold and extrapolates forward. If 14 million monthly visits are already gone — and the Rock Health and Gallup data suggest the number is growing — the projection's baseline is wrong. The shortage may be smaller than estimated, or may be in different specialties than estimated, or may reverse entirely in some categories. We don't know, because the data that would tell us doesn't exist in the places we look for it.

This is the $50B blind spot, updated. The 2008–2016 version was a decline in primary care visits that claims-based analyses could not fully measure. The 2025–2026 version is a decline in everything AI can plausibly substitute for: symptom interpretation, medication questions, follow-up clarification, diagnosis research. Gallup found that patients are using AI for all of these.

How big is 14 million, really?

The U.S. generates roughly 1.07 billion physician office visits per year, according to the most recent NAMCS data — about 89 million per month. Gallup's 14 million represents about 16% of monthly visits, displaced by AI.

That's the context for the 14 million number. It isn't a niche phenomenon. It is a sixth of current monthly physician office volume, gone, already, with no claims data to show for it.

The clinician view is catching up

When I posted the Gallup data on X last Tuesday, the response that's stayed with me came from Dr. Martha Gulati, a preventive cardiologist at Cedars-Sinai: "This is remarkable but also true. This is why AI needs to be validated and the medical community needs to be a part of this to ensure we are meeting patients where they are. Which is on these apps, SoMe and the internet."

That's the right instinct, and it's worth taking seriously. But it also touches on the hardest operational problem in this whole picture. Validation, as the medical community has traditionally practiced it, is a process that takes years. The tools patients are using change weekly if not daily. ChatGPT in April 2026 is a meaningfully different product than it was in October 2025. By the time a rigorous validation study on a specific chatbot is designed, funded, run, peer-reviewed, and published, the product has shipped a dozen new versions.

The medical community's instinct is to slow things down until the evidence is in — but the patients aren't waiting. Whatever validation looks like in this environment, it will have to happen at consumer-software speed — or it won't happen at all.

What health system strategy should do about it

Two things.

First, stop treating projected visit volume as a reliable base for capital and capacity planning. It is not a reliable base. It is the output of a model that misses a significant and growing share of what patients are doing.

Second, improve the model by measuring that share (i.e. what patients are doing outside the system). This is harder, but by no means impossible. Patient-reported utilization surveys, employer health plan data, and direct consumer research can fill some of the gap that claims data cannot. If your strategy is relying on claims data alone, your strategy is flying blind on the most important variable.

In a nationally representative survey of 5,500 U.S. adults released April 15, 25% reported using AI for health information or advice. Among recent users, 14% said the AI-generated advice led them to skip a provider visit in the past 30 days. Projected to the adult population, that's roughly 14 million missed appointments in a single month.

The U.S. generates about 89 million physician office visits per month. Gallup's 14 million represents about 16% of them — displaced by AI, already, in 2026.

Two key details:

First, only 4% of AI users strongly trust the accuracy of the health information they're getting. The skipped visits aren't happening because patients believe AI is better than their doctor. They're happening because the alternative is worse: 14% of recent users cited inability to pay, 16% cited lack of access, and 21% cited feeling dismissed by a provider in the past. AI is filling a gap the delivery system created.

Second, this is invisible in claims data. A visit that never happens generates no CPT code and no line item. It shows up nowhere in the revenue cycle, nowhere in board dashboards, nowhere in the physician shortage projections that assume current utilization patterns hold. Ganguli et al. documented the same pattern for the 2008–2016 period — a 24% decline in primary care visits, roughly three-quarters of it outside the claims data entirely. The Gallup survey is that phenomenon, updated and accelerated.

The operating implication for health system strategy is straightforward: if you are planning capacity, capital, or workforce against projected visit volume, your plans are wrong.

A longer piece on this is coming Monday.

Eighteen percent said they adjusted a medication.

Not researched a medication. Adjusted a medication: changed dose, stopped, started. Among the roughly one in three adults now using AI chatbots for health, nearly one in five has changed their own prescription based on a chat window.

Where the decision happened

Seventy-three percent of those AI users went to ChatGPT. Five percent used a chatbot offered by their provider. Four percent used a payer's.

One place this prescription medication adjustment didn’t happen: a patient portal or a nurse line. Or a clinic. Ninety percent happened in a window the health system doesn't operate, doesn't see, and can't review. Whatever the chatbot said, whatever the patient thought it said, whatever they decided to do — none of it is known to the institution that wrote the original prescription.

It’s not a one-time thing

Sixty-four percent of AI health users engage weekly or more. About one in five do it daily. The 18% medication-adjustment figure is describing a behavior embedded in an ongoing consultation pattern — one where the consultant is a chatbot and the original prescriber finds out at the next visit, if there is one.

The chart has no field for this

The EHR shows the regimen the clinician prescribed. The patient is on something else. Medication reconciliation at the next encounter is supposed to catch the gap, and sometimes it might. Ernst and colleagues, working before consumers had a clinical advisor in their pocket, found discrepancies in 26.3% of charts at the point of prescription renewal: the EHR didn't match what the patient was doing. That was the baseline.

Now that baseline is changing, and not for the better. Every additional channel through which a patient can get a plausible-sounding answer about their medication is another source of off-record adjustment. Eighteen percent of AI users acting on chatbot input is one such channel, and it's the biggest one yet.

What I missed last week

Last week in The $50B Measurement Blind Spot I argued that demand migration (from healthcare to consumer tech) shows up as visits the claims data can't see — going back to the Ganguli finding that primary care visits fell 24% from 2008 to 2016, with three-quarters of the decline invisible to standard utilization measures.

But it turns out that even inside the healthcare visits that do still happen, a meaningful share of what gets subsequently decided is being decided somewhere other than the clinic or ED or urgent care. The divergence between the prescribed regimen and the actual regimen is increasing, and the divergence is happening through a channel that produces no claim, no portal message, no chart entry.

The strategic question for a health system isn't whether AI chatbots will eventually affect prescribing behavior. It's how much of the prescription record on the books today is already fictional. And what does that mean for healthcare.

The one thing they’re not tracking at all? The demand leaving the system entirely. 

A finding that should have set off alarms

In 2020, Harvard researcher Ishani Ganguli and colleagues published data in the Annals of Internal Medicine that should have reshaped healthcare strategy: from 2008 to 2016, primary care visits among commercially insured adults dropped 24%, with adults 18–34 showing the steepest decline at 29%. Preliminary data presented at AcademyHealth 2023 Annual Research Meeting suggest the decline continued through at least 2021.

Now imagine Ford discovering car sales had dropped 24%. Every analyst would track monthly figures. Every dealership would know. The board would have real-time dashboards.

Healthcare’s response? There was none. Recently, at a conference of hospital executives and board members, I asked the room if they were aware of this research. Blank stares. No one had seen the data.

Where’d they go?

The researchers tried to track where patients went instead of to primary care. What they found:

Alternative venues: Urgent care, retail clinics, ED, telemedicine. Still provider-mediated, still billed to insurance, still visible in claims. This accounted for about one quarter of the decline.

Everything else: Three quarters of the decline didn’t show up in claims at all — leaving the researchers unable to determine from claims data alone where that demand went. It could include three very different things:

1. Forgone care — people skipping treatment because of cost or access barriers. The authors pointed out that copays for problem-based visits increased 32% over the study period, and the visit decline was steepest in lower-income areas—suggesting cost might indeed be a factor.

2. Self-pay provider care (Hims, Ro, cash-pay telehealth) — still a clinician, but invisible to insurance.

3. Provider-less care — symptom checkers, WebMD, Apple Watch, ChatGPT. The authors noted that patients (especially younger patients) “accustomed to convenient, on-demand services in other aspects of their lives” were increasingly comfortable with “Internet-based health information, symptom checkers, and virtual patient communities” that “may preclude seeking formal medical care altogether.”

Although we first learned of this puzzle six years ago, we still don’t know the breakdown among these three categories. And we’re not even trying to find out.

The fifty-billion-dollar question

Primary care spending in the U.S. runs roughly $300 billion a year. A 24% decline represents over $70 billion in missing visits—and three quarters of that is unaccounted for. That’s a fifty-billion-dollar question: where did all that demand go?

If people are forgoing care entirely because of cost or access barriers, that’s a coverage, affordability, and capacity problem. Health systems might respond with sliding-scale pricing, expanded charity care, advocacy for better insurance coverage, or investments in scheduling capacity and workforce.

If they’re shifting to self-pay provider care—companies like Hims and Ro, which can generate billions in revenue on subscription models outside traditional reimbursement—that’s competition.

If they’re moving to provider-less care—handling health questions with ChatGPT, monitoring their hearts with Apple Watches, managing anxiety with apps—that’s a different threat entirely. That demand may never come back. No pricing strategy or convenient hours will recapture someone who’s decided they don’t need a clinician at all.

Each scenario has massive implications—for health systems planning workforce and capital investments, and for policymakers setting coverage and cost-sharing rules.

The measurement blind spot

Researchers have studied pieces of this puzzle—DTC telehealth utilization, consumer health tech adoption, cost-driven foregone care. But studies show DTC often drives new utilization (88% in one analysis) rather than direct substitution, and recent reviews highlight gaps in understanding system-wide impacts. No one has connected the dots to answer the strategic question: of the $50 billion in missing primary care demand, how much is going where?

How many diagnoses did Apple Watches make last year? Apple knows. Kaiser doesn’t. How many people skipped a primary care visit because of a $50 copay? How many used Hims instead of their PCP?

The data to answer these questions likely already exists—in surveys, in tech platforms, in self-pay provider records. But no one in healthcare is combining it, analyzing it, or acting on it.

This is the measurement blind spot at the center of healthcare strategy. Executives obsessively monitor demand shifting within their systems. What they don’t track, at all, is demand that leaves the system—or why.

Time to start measuring

We don’t yet know the split among cost barriers, self-pay competition, and consumer tech—but we need to find out. And right now, it’s not anyone’s job.

Each of those three possibilities demands a different response — and without knowing which one dominates, no rational response is possible. Every health system CEO should be asking: Of the demand that’s left your system, how much was lost to cost barriers, how much to self-pay competitors, and how much to provider-less care? If you can’t even estimate the split, you’re making billion-dollar capital allocation decisions blind.

Answering this will require new data sources—patient surveys, DTC partnerships, self-pay records, consumer trend analysis. But healthcare ignores this sea change at its peril. The organizations that start measuring now, and building strategy around what they find, are much more likely to be the ones still standing in 2035.

I joined Scott Becker this week on the Becker's Healthcare Podcast for a wide-ranging conversation about why healthcare is getting better and getting smaller.

We covered the 24% decline in primary care visits that nobody in healthcare seems to know about, why health care spending growth has been falling for fifty years, GLP-1s as demand elimination rather than disruption, what autonomous vehicles will do to trauma revenue, and why health systems are tracking the wrong threats.

Scott was a generous host and asked the right questions. Worth a listen if you want the short version of the thesis.

Listen here.

David Oks recently published a brilliant essay dismantling one of economics' favorite bedtime stories: the parable of the bank teller and the ATM. The standard version goes like this: ATMs arrived, everyone predicted mass unemployment for tellers, and yet teller employment rose. Technology doesn't destroy jobs. Relax.

J.D. Vance cited this story in a New York Times interview. So have Daron Acemoglu, David Autor, Matt Yglesias, and Eric Schmidt. It became, as Oks puts it, a load-bearing parable for anyone arguing that technology creates more jobs than it destroys.

And through about 2010, the story was true. ATMs reduced the number of tellers per branch — from about 21 to 13 — but that made branches cheaper to operate. Banks opened more branches. Total teller employment held steady or even grew.

Then the iPhone showed up. And between 2010 and 2022, bank teller employment was cut roughly in half.

Task automation vs. paradigm replacement

Oks draws a clean distinction. The ATM tried to do the teller's job — handle cash, process deposits — faster and cheaper. It automated tasks within the existing paradigm. In Clayton Christensen's framework, the ATM was a sustaining technology for bank branches: it let them do what they were already doing more efficiently and less expensively. The customer still came to the branch. The branch still existed. The teller's role shifted from handling cash to selling products. And per Jevons's Paradox, as it got cheaper to run a branch, banks opened more of them.

But the iPhone didn't let the branch do its work more efficiently. It made the branch unnecessary. Mobile check deposit. Venmo. Real-time balance checks at 2 a.m. The iPhone created an entirely new paradigm in which the old one simply didn't need to exist.

Oks' key insight: it is paradigm replacement, not task automation, that actually displaces workers.

The real story is more complicated — in banking and in healthcare

Oks' essay got substantial pushback. Critics pointed out that the ATM/iPhone binary is too clean. Online banking was used by 44% of Americans, and UK bank branches were closing, by 2005, well before the first iPhone shipped. The 2008 financial crisis cratered the banking industry and is a major confounder in the teller employment data. And even the ATM-era "complementarity" story — the claim that ATMs complemented tellers rather than replacing them — was partly an illusion: ATMs did eliminate a third of teller jobs per branch, but deregulation-driven branch expansion masked that decline.

The real story is a constellation of digital tools — online banking, then mobile apps, then Venmo, then neobanks — that gradually made the branch less central, until one day it wasn't central at all. No single technology delivered the killing blow. The paradigm just gradually eroded.

The critics are right — and their more complicated version is a better match to healthcare than Oks' binary.

Healthcare's ATMs

Healthcare is living through the ATM phase right now — and telling itself the same reassuring story the banking industry told in 2005.

The "ATMs" of healthcare are clinical AI tools. Ambient scribes that write your notes. AI co-pilots that suggest differential diagnoses. Tools like OpenEvidence, DoxGPT, and Glass Health are genuine innovations. They make the doctor faster, reduce burnout, maybe even improve accuracy.

But notice what they don't do: they don't question whether you need to see the doctor in the first place. They automate tasks within the existing paradigm. The patient still goes to the clinic. The doctor still renders a judgment. The visit still gets billed. The teller's job description changes, but the teller survives — because the branch is still open.

This is what Oks calls "fitting capital into a labor-shaped hole." Health systems are taking a powerful new technology and cramming it into workflows designed around a human in an exam room. The result is a faster, cheaper version of the same encounter — just as the ATM produced a faster, cheaper version of the same branch visit.

And just as branch expansion masked the early decline in per-branch teller employment, volume growth from aging boomers and expanded access mandates may be masking a similar dynamic in healthcare — clinical AI reducing the need for certain staff roles, invisible for now in the aggregate numbers. The complementarity story looks great on paper, right up until the moment the underlying demand shifts.

Healthcare's iPhone

Now look at what's happening outside the exam room.

An Apple Watch detects atrial fibrillation and alerts the wearer before they ever call a cardiologist. A consumer orders a lipid panel from a DTC lab, gets results on their phone, and adjusts their statin dose with an AI health assistant — no appointment necessary. A continuous glucose monitor coaches a pre-diabetic patient through dietary changes that prevent the diabetes diagnosis entirely. GLP-1 drugs eliminate the obesity that would have generated decades of cardiovascular, orthopedic, and endocrine visits. Early signals are visible: between 2022 and 2023, GLP-1 prescriptions for obesity more than doubled while bariatric surgery rates fell 25.6%. Some hospital bariatric programs have already closed.

None of these are trying to make the doctor visit better. They're making the doctor visit unnecessary.

And just like banking, there is no single "iPhone." There's a constellation of technologies — wearables, DTC labs, AI health assistants, GLP-1s, telehealth, and, yes, the iPhone itself — each removing one more reason to visit the doctor, none dramatic enough on its own to trigger alarm. The disruption is diffuse, which is precisely why it's invisible to the people running the system being disrupted.

Where the analogy breaks

Here's where the banking analogy breaks — and why healthcare is in a worse position than banking was.

When mobile banking killed branches and teller jobs, the banks captured much of the transition. JPMorgan Chase built the app. Citibank built the app. The technologies might have disrupted branches but they made the overall business of banking more profitable than ever — it just needs fewer buildings and fewer humans. The workers lost, but the corporations won.

Healthcare isn't getting that deal. The health tasks migrating out of the exam room aren't migrating to apps from Kaiser or Mass General or any other healthcare system. They're migrating to Apple, Whoop, Eli Lilly, and a growing ecosystem of DTC companies that the health system doesn't own, doesn't operate, and can't bill for. And some of the demand — the conditions that GLP-1s prevent, the injuries that autonomous vehicles will never cause — isn't migrating at all. It's simply disappearing.

In banking, the iPhone threatened workers and branches. In healthcare, the constellation threatens workers, facilities, andthe business model itself — because the value is being captured outside the organizations that have historically delivered care.

None of this will happen overnight. Healthcare has regulatory moats, liability exposure, and reimbursement plumbing that banking never had to navigate. The migration will be slower and messier than banking's was. But it is happening.

What the board should be asking

Oks' essay is about employment — does technology kill jobs? The ATM parable said no. The iPhone proved yes, just not the technology everyone was watching.

Applied to healthcare, the same logic holds. Clinical AI tools won't kill healthcare jobs (as the ATM didn't kill banking jobs). But the technologies replacing the doctor visit — wearables, DTC labs, AI health assistants, GLP-1s — will reduce how many we need (as mobile and online banking did). My prediction is that in ten years we will need fewer physician visits per population, not more. Fewer hospitals, not more. And fewer doctors as currently constituted, not more.

Those are the employment questions, and they track Oks precisely. But as I've argued, the employment problem is only half the story — healthcare isn't just losing workers, it's losing the demand itself, to entities it doesn't own.

Most health system boards are asking the ATM question: which AI tools should we buy to make our operations more efficient?

Almost none are asking the iPhone question: does the visit itself have a future — and if it doesn't, can we capture the value wherever it goes?

Banking answered that second question by building mobile banking apps. Healthcare hasn't started to ask it — while others are already building.

I've been writing about the forces reshaping healthcare — what I call the Five Migrations and demand elimination — for the past year. If this framework resonates, you might start with "The Economist Wrote About Healthcare Disruption — But Missed the Disruption" or "When Health AI Makes a Mistake, Who's Liable?"

In a classic AI-is-weird way, the models answered anyway — confidently, in detail, with elaborate reasoning about images they never saw. The kicker? On some radiology tests, the models scored up to 99% as well without the image as with it. A small text-only model, trained on radiology Q&A with images stripped out, outperformed radiologists by more than 10% on a chest X-ray test. Without looking at any images.

One implication: Most of the commentary has framed this as an argument against deploying radiology AI. The paper's own conclusion is narrower: the standardized tests are broken, and we need better ways to measure whether AI is actually using visual information. Makes sense.

But I think the data is also asking a more provocative question — one the authors don't raise directly. How often is the diagnosis already in the EHR before anyone ever looks at the image — making the imaging unnecessary for diagnostic purposes?

Experienced clinicians will tell you they frequently know the probable finding before the film comes back. 62-year-old smoker with chronic cough and weight loss? You have a strong prior. The imaging confirms it. What the Stanford team demonstrated — accidentally — is that an AI can formalize those priors so reliably it tops a radiology test blind.

This is very speculative on my part: the paper doesn't claim imaging is unnecessary, and neither do I. But if a text-only model can outperform radiologists on a visual test using nothing but clinical context, it's worth asking: what portion of imaging volume is confirmatory rather than diagnostic — a clinical ritual more than a clinical necessity? And confirmatory rituals are exactly the kind of that eventually just disappears.

The tests do need fixing — the scorecards should distinguish between visual grounding and text inference. (As I wrote in "We Want to Grade the AI. Did We Grade the Doctor?" — we have a long history of not testing rigorously enough.) But the deeper story may not be that the AI cheated. It may be that the clinical information was sufficient all along, and the imaging was unnecessary.

One implication: Radiologists have spent nine years celebrating the "failure" of Nobel prizewinning AI scientist Geoffrey Hinton's famous 2016 prediction that we should stop training radiologists because we already have enough to last until the machines take over. Now the CEO of America's largest public hospital system says he's ready to let the machines take over — at least for first image readings. Katz isn't talking about firing radiologists tomorrow. He's talking about letting AI handle first reads and having radiologists review the flags. And he's not a researcher, he's a hospital CEO trying to solve a staffing and cost problem, asking his peers on a public stage why they aren't pushing regulators to let it happen.

The radiology community's response has been predictable: another confidently uninformed administrator being duped by AI hype. Maybe. But Katz runs eleven hospitals. Scott runs a small system where every dollar matters. These aren't people chasing novelty. They're people doing math.

Hinton recently acknowledged he was wrong on timing. But the endpoint he suggested — machines taking over image reading — just got boosted by three hospital CEOs on a public stage.

Here's what's remarkable about those numbers: no one has adequately evaluated whether OpenEvidence is safe. A systematic review submitted to npj Digital Medicine, currently in preprint, put it bluntly: "The rapid adoption of OpenEvidence among clinicians outpaces the sparse research available."

A tool used daily by 430,000 doctors, across 10,000 hospitals, in a country that spent $4.9 trillion on healthcare last year — and the evidence base for it wouldn't survive a journal club.

A crowded market with an empty scoreboard

OpenEvidence isn't the only company offering AI-powered clinical decision support to doctors:

• Doximity — whose platform reaches 85% of U.S. physicians, and whose HIPAA-compliant text messaging system I frequently rely on — offers DoxGPT, which had 300,000 unique clinician users in just the last three months of 2025.

• Glass Health, backed by Y Combinator, offers AI-generated differential diagnoses and clinical plans from patient summaries.

• Vera Health offers AI-powered clinical answers with direct links to peer-reviewed sources.

• UpToDate — long the gold standard for clinical decision support — launched its generative AI product, Expert AI, in September 2025.

• DynaMedex, iatroX, and others are crowding in behind them.

Every one of these platforms claims to be evidence-based. Every one of them cites partnerships with prestigious journals and institutions. Every one of them says it's built for safety.

Not one of them appears to have ever been independently tested against the others. I was unable to find any study — not one — that takes OpenEvidence, DoxGPT, Glass Health, Vera Health, and even two or three of their competitors, gives them all the same panel of clinical cases, and tells doctors which one performs best.

What's worse: as we'll see, even the most widely used of these tools has only the thinnest evidence base for accuracy — and the rest appear to have none at all.

How good is each medical AI? We don't know.

For OpenEvidence — the most widely used of these tools — the answer is barely. The npj Digital Medicine preprint found that existing studies were small, narrow, and confined to specific clinical domains. Accuracy varied in complex scenarios. The platform often reinforced rather than altered clinical decisions — which means we don't really know whether it's helping doctors get the right answer, or just telling them what they already think.

For DoxGPT, Glass Health, Vera Health, and the rest? I was unable to find any comparable body of independent evaluation. Even UpToDate — the most trusted name in clinical decision support for three decades — has not published accuracy benchmarks for its new Expert AI product. When asked about accuracy by Newsweek, Wolters Kluwer's chief medical officer said it's been "pretty good." A spokesperson added: "We don't have shareable benchmarks yet." The product is already available to 250,000 clinicians.

A Sermo survey of physicians found that 44% cited accuracy and misinformation risk as their top concern about tools like OpenEvidence. Another 37% said peer-reviewed validation was the single factor that would most encourage adoption.

Doctors are asking for the evidence. Nobody is producing it. But 40% of doctors are still using it — including me sometimes, I'll admit, though probably less after researching this post.

Which medical AI is best? We don't know that, either.

The closest thing to a head-to-head comparison: Doximity published a side-by-side evaluation of its own DoxGPT against OpenEvidence, the traditional UpToDate online reference (not the newer Expert AI), and ChatGPT. More than 1,000 physicians evaluated the answers, and DoxGPT was selected as the "best clinical answer" at more than twice the rate of OpenEvidence. But that study asked physicians to pick which answer they preferred — with no independent gold standard, no expert panel adjudicating correctness, no measurement of whether either answer would have led to a safe outcome. Preference and accuracy are different things. It was also Doximity's own study, conducted on its own platform.

Doximity has since launched PeerCheck, co-chaired by Eric Topol and former Surgeon General Regina Benjamin, to bring physician peer review to DoxGPT outputs. That's a step in the right direction — but it's still one company's product, reviewed by that company's advisors.

The Hugging Face Open Medical-LLM Leaderboard at least attempted to maintain a running, independent comparison of models on standardized medical benchmarks. It has since gone dormant — a casualty of the broader Open LLM Leaderboard's retirement. The benchmarks, Hugging Face acknowledged, were becoming obsolete.

That's it. That's the evidence base for the tools doctors are using to make clinical decisions right now.

Why this is allowed

Imagine a pharmaceutical company told you its new cardiac medication was used by 40% of cardiologists and had never been tested in a large-scale, independent trial — just a handful of small retrospective studies, most of which found the drug "reinforced" existing treatment choices rather than improving them.

The FDA would never approve it. But that's exactly where we are with medical AI. And the regulatory framework makes it possible.

Under the 21st Century Cures Act, clinical decision support software is exempt from FDA regulation — provided it allows the physician to "independently review the basis" of its recommendations. In January 2026, the FDA further loosened this, clarifying that even software providing a sole medical recommendation can be exempt, as long as the clinician can evaluate the underlying logic.

In practice, this means that the AI platform your doctor consulted before writing your prescription went through less regulatory scrutiny than the prescription itself.

What needs to happen

This isn't an argument against medical LLMs. These tools are very likely better than a physician relying on memory, a five-year-old UpToDate article, or a hurried Google search between patients.

But "very likely better" is not a standard we accept anywhere else in medicine. We need:

1. An independent, regularly updated benchmark that tests OpenEvidence, DoxGPT, Glass Health, Vera Health, and their competitors on the same clinical case panels, scored by the same physician panels, on safety and accuracy alike.

2. Head-to-head trials with clinical endpoints. Not just "did the AI give a correct answer?" but "did patients whose doctors used Platform A have different outcomes than patients whose doctors used Platform B?"

3. Transparent reporting of safety data. If Doximity can mobilize 10,000 physician experts for PeerCheck, an independent body can do the same — across platforms.

These tools are reshaping how medicine is practiced in America right now. The question isn't whether AI should be in the exam room — it already is. The question is whether we're going to evaluate it like a medical intervention, or keep treating it like Microsoft Word.

For a profession built on evidence-based medicine, the answer should be obvious.

The article's central premise goes like this: American healthcare has "proved resistant to treatment from disrupters." Google tried and failed. IBM tried and failed. Amazon tried and failed. But now — with AI improving, patients fed up, and regulators loosening — maybe this new wave will finally break through.

That premise is wrong. Not the part about the new wave — those companies are real, and some of them will matter. The part that's wrong is the claim that past disruption failed. It didn't fail. It succeeded — massively, over decades — and the healthcare system (and the Economist) never noticed, because the disruption happened outside the walls, and no one was measuring.

Decades of disruption, ignored

The Economist lists Google as its first example of a failed disrupter. In 2008, Google launched a platform to help people store and manage their health records. It shut down in 2011 after failing to gain users.

Fair enough. Google Health the product was a flop. But Google Search has been arguably the most successful healthcare disrupter in history. How many people have Googled a symptom over the past twenty years and decided they didn't need a doctor? How many sore throats, rashes, back pains, and anxious late-night questions were resolved — well or poorly — by a search engine instead of a physician?

No one knows. Because the healthcare system only measures health activity that happens in clinics, or wears a white coat. Google Search didn't show up in any claims database or utilization report. It generated no encounters, no referrals, no billing codes. So the industry filed Google under "failed disrupter" and moved on — while the actual disruption continued at a scale no startup has come close to matching.

Google isn't even the oldest example. The disruption started in the pharmacy aisle. Every time the FDA switches a drug from prescription to over-the-counter — Prilosec, Claritin, Flonase, Plan B, Narcan — a healthcare task migrates from the exam room to the consumer. The patient who once needed an appointment, a copay, a provider, and a prescription now walks into CVS and handles it themselves. That visit disappears from the system. No claim is filed. No encounter is recorded.

OTC switches have been happening for decades. They are explicitly designed to remove doctors from the equation, and that is what they have been doing, very effectively. Nobody calls them disruption. But they are, and one of the most effective forms of it — just invisible to the institutions losing the volume.

The Data Shows Healthcare Has Been Shrinking for Years

If past disruption really failed — if healthcare truly "proved resistant" — you'd expect the data to show a system holding steady or growing. It shows the opposite.

Real per capita healthcare spending growth in the United States has been falling for half a century (per CMS data). In the 1960s, per capita expenditure (in 2023 dollars) grew 83% over the previous decade. By the 1980s: 51%. The 2000s: 30%. The most recent decade: 28%. And since 2000, real per capita spending has actually declined in absolute terms multiple times — something that has never happened before in the 60+ years or so we've been keeping track.

The Economist opens with a striking number: America spends $5 trillion a year on healthcare. That figure is meant to convey the enormity of the fortress. But on a per-person, inflation-adjusted basis, the fortress has been losing ground for decades, and has outright contracted in several recent years. Something has been siphoning demand out of the system for a long time. Whatever that is, healthcare doesn't look like a system that "proved resistant to treatment from disrupters."

Then there are the visits. A 2021 study by Ganguli and colleagues in JAMA Internal Medicine found that primary care visits in the United States fell 24% between 2008 and 2016. Twenty-four percent! And three-quarters of that decline couldn't be accounted for by any shift visible in claims data — no transfer to urgent care, no movement to telehealth, no substitution by specialists. The visits just vanished — though the study authors suggested that a large portion could be attributed to former patients, especially "younger, healthier adults, who may have fewer routine care needs and be increasingly comfortable with online self-care."

Slowing spending growth. Absolute spending declines. A quarter of primary care visits gone with no trace in claims data. Decades of OTC switches moving treatment tasks to pharmacy aisles. Billions of health queries answered by Google.

This isn't a system that resisted disruption. This is a system that was disrupted and didn't notice.

The New Wave Is Real — and It Fits the Pattern

The Economist is right that something new is happening. But the new companies aren't breaking through a wall that previous disrupters couldn't breach. They're accelerating migrations that have been under way for a generation.

I've been writing about this as the Five Migrations — the systematic movement of healthcare tasks from clinical settings to consumers. Knowledge. Diagnosis. Treatment. Monitoring. Interpretation. Each follows the same arc: a task that once required a clinician in a clinical setting becomes something a patient can do themselves.

The Economist's company catalog maps directly onto this framework:

Knowledge and Interpretation. OpenAI says more than a quarter of ChatGPT's 900 million weekly users ask at least one health question — most outside clinic hours. This is the Knowledge Migration that started with Google Search, now turbocharged by AI that can hold context, interpret lab results, and personalize. ChatGPT didn't create the demand. It's serving demand that has been migrating away from clinicians for twenty years.

Treatment. Hims & Hers and Ro started with hair loss and erectile dysfunction — low-stakes categories that most physicians barely contested. Now they sell GLP-1 prescriptions for weight loss, one of the most medically significant treatment categories in a generation. That's the classic Christensen pattern: disrupters enter at the bottom of the market, where incumbents don't bother to compete, then move upmarket. Drugmakers are already responding by selling directly to consumers. Every one of these transactions is a visit that didn't happen — the same Treatment Migration that OTC switches have been driving for decades, now extending into categories that matter enormously to the traditional system.

Diagnosis and Monitoring. Prenuvo sells full-body MRI scans for around $1,000. Function Health sells comprehensive blood panels for periodic monitoring. These look new, but the Diagnosis Migration started decades ago — with the home pregnancy test. Then came home HIV tests, home strep tests, home COVID tests. Each one moved a diagnostic task from a clinician's hands to a consumer's. The Monitoring Migration followed the same path: the home blood pressure cuff, the home glucose monitor, Fitbit, Apple Watch, continuous glucose monitors now worn by people without diabetes. Prenuvo and Function Health are the latest step, not the first. And like their predecessors, they're largely serving demand the clinical system ignored — because guidelines didn't call for screening, or because the patient never made it to the appointment where the conversation would have happened.

The Gatekeeper Objection

The article surfaces the standard physician rebuttal: consumer-purchased scans and tests will flag benign conditions and strain an already stretched system.

This is the exact argument incumbents make at every stage of a Christensen disruption. The new product is inferior. It creates waste. It serves people who don't really need it. Minicomputers were toys. Online brokerages were for amateurs. Nurse practitioners shouldn't practice independently.

The pattern: gatekeepers define what counts as necessary, then object when consumers make a different assessment. But the consumer who pays $1,000 for a Prenuvo MRI has decided the information has value to them — regardless of whether a physician would have ordered it.

And most of these companies aren't competing with doctors. They're competing with nothing — which is what many people were getting before. When a third of American adults are skipping or delaying care because of cost, doing nothing ("nonconsumption" in Christensen terms) is the real alternative. The startups are competing, and winning, against that.

What you can’t see CAN hurt you

The Economist concludes that American healthcare "may not be immune to change for much longer."

That gets the tense wrong.

Per capita spending growth has been falling for fifty years. Real per capita spending has declined outright, multiple times, since 2000. Primary care visits dropped 24% before Hims existed. Billions of health questions were answered by Google before ChatGPT launched. OTC switches have been migrating treatment tasks out of the exam room for generations.

The reason the Economist — and most of the healthcare industry — keeps missing this is structural. The system measures what happens inside its walls. Claims data. Encounter records. Utilization reports. These instruments are precise, comprehensive, and blind to the most important trend in healthcare: the growing share of care that happens outside the system entirely. The spending data confirms it. The visit data confirms it. The companies the Economist profiles confirm it.

So The Economist noticed the disruption of the last three years — it just missed the last three decades.

Previously: We Want to Grade the AI. Did We Grade the Doctor?

Coming next: Your Doctor's Favorite AI Has Never Been Tested

The reaction has been swift. The AMA warned that accuracy claims don't replace clinical judgment. The American College of Physicians said prescriptions shouldn't be issued without physician involvement.

Most recently, Hopkins trauma surgeon Joseph Sakran and medical student Rahul Gorijavolu‍ ‍wrote in the Washington Post that the program's safety case rests on a single unreviewed preprint — authored entirely by equity holders, with underlying data unavailable to outside researchers.

They're right about that. Doctronic's evidence is thin. The study hasn't been peer-reviewed. The data hasn't been shared. The 99.2% accuracy figure everyone cites was derived from urgent care encounters, not the chronic medication renewals Doctronic is actually doing in Utah.

All fair criticisms. But here's the question nobody is asking:

What's the evidence base for the process Doctronic is replacing?

Prescription renewals are a big part of everyday medicine. In many practices, they move through a fast, pieced-together workflow: a pharmacy request, a portal message, a medical assistant or nurse using a protocol, a chart routed to a covering clinician, a physician clicking "approve" between visits. The details vary from place to place, but the overall picture is familiar. This is high-volume outpatient work done under time pressure.

Has that process been studied? Yes, but not in any single decisive way. There is no landmark trial that established the ordinary human prescription-renewal process as a gold standard. There is no widely accepted national error rate for physician refill approval. What exists instead is a scattered literature on refill workflow, medication discrepancies, prescribing and monitoring errors, pharmacist interventions, and refill protocols designed to catch the things ordinary practice often misses.

A few examples:

• How Physicians Renew Electronic Prescriptions in Primary Care: Therapeutic Decision or Technical Task — a direct look at renewal practice, and a useful reminder that renewal is clinically important but not deeply studied as its own process.

• Medication discrepancies in an outpatient electronic medical record — found discrepancies in 26.3% of charts, often because the medication list in the record was incomplete.

• The prevalence and nature of prescribing and monitoring errors in English general practice— found errors in 4.9% of prescription items, with missed monitoring a major issue.

• Implementation of a Multidisciplinary Medication Refill Protocol — shows why systems build protocols in the first place: refill work is large, time-consuming, and benefits from structured checks and escalation.

• Prescription renewal request reviews by student pharmacists — another demonstration that many "routine" renewals turn out to need intervention, optimization, or follow-up.

• Comparison of interventions in a pharmacist-managed refill model to usual physician care — the closest thing to a head-to-head: when pharmacists reviewed refill requests carefully, they found issues with 7.0% of refills. Usual physician care caught 3.5%.

What that literature shows is that the standard renewal process has a lot of chances to miss things: missing labs, outdated medication lists, overdue follow-up, drug interactions, medications that should have been adjusted, and medications that should have been stopped. It is a human process that usually works, sometimes misses important things, and has been made safer over time with protocols, pharmacy review, and extra layers of checking.

That matters because the case against AI prescription renewal is often argued as though the human alternative is already a proven standard. It is not. It is the existing standard — which is a different thing.

The Washington Post op-ed also points, reasonably, to the possibility that some medical AI tools may perform differently across demographic groups. Fine. That should be examined. But notice the asymmetry. When AI enters medicine, we ask for subgroup analysis, prospective evidence, peer-reviewed validation, and proof that it works. We rarely ask for the same level of evidence about the human workflow it is being compared against.

That is the real point. The standard human renewal process is not some rigorously validated benchmark. It is an old, familiar process, used at enormous scale, with known weak points and a thinner evidence base than most people would assume for something so common and so important.

This is the double standard at the center of the AI-in-healthcare debate right now. We are applying a level of evidentiary scrutiny to AI that we have never applied — and show no signs of applying — to the human processes AI is replacing.

I'm not arguing that Doctronic should get a pass. It shouldn't. The evidence should be better. The preprint should be peer-reviewed. The data should be available. The study population should match the deployment population.

But if the standard is "rigorous, independent, peer-reviewed evidence of safety before deployment" — and it should be — then that standard has to apply to both sides of the comparison. You can't demand a randomized controlled trial from the AI while grandfathering in the physician on the basis of tradition.

The authors of the WaPo op-ed write that "the window to act is this year, before autonomous AI prescribing expands." They're right about the urgency. But the most useful thing we could do with that window isn't just to test the AI. It's to test both — and find out what we've actually been living with.

This piece focuses on liability in enterprise and clinical AI. For the consumer side of the same question, see When Consumer Health AI Makes a Mistake, Who’s Liable?.

The Usual Models

Physicians have always used tools, and those tools have always had a sensible liability structure — though not a single one. It depends on what kind of tool you're talking about.

Active tools

For tools that perform an action on the physician’s behalf — lab tests, diagnostic devices, stain reagents — liability follows the source of the error. When a pathologist misreads a correctly performed biopsy stain, the pathologist bears the liability. When the stain reagent is defective — contaminated, mislabeled, degraded — the manufacturer bears it. If a physician prescribes the wrong medication because a lab test was faulty, the liability lies with the lab. If a physician prescribes the wrong medication despite an accurate lab result, the liability lies with the physician. The responsible party is the one whose failure caused the harm.

Passive tools

Reference materials work differently. If a physician follows incorrect guidance from a textbook or clinical reference and a patient is harmed, the liability lies with the physician — not the publisher. Publishers have long disclaimed responsibility for clinical decisions made on the basis of their content, and courts have consistently upheld that position. The physician is expected to exercise independent judgment; the book is just a source.

These two models have coexisted in medicine for decades. The distinction matters because AI sits at the boundary between them — and how you classify it determines who pays when something goes wrong.

How AI Departs from Both Models

Most AI tools used in clinical settings superficially resemble active tools — lab tests, diagnostic devices, things that perform an action and return a result. But their contract terms put them squarely in the passive category. And then they go further.

The traditional liability model would suggest: if the physician uses the AI’s information incorrectly, it is the physician’s fault — but if the AI provides incorrect information (a hallucination — fabricated or incorrect output presented with apparent confidence), the AI vendor bears the liability. That is not the approach these contracts take. If the physician makes the mistake, it is the physician’s fault. If the AI makes the mistake, it is still the physician’s fault. Unlike the defective stain reagent manufacturer or the lab that ran a faulty test, the AI vendor has contracted away responsibility even for its own product’s failures.

OpenAI’s terms for ChatGPT for Healthcare — a product explicitly marketed for clinical workflows, with HIPAA compliance and BAAs (Business Associate Agreements) — tell clinicians they “should always verify the information provided” and cannot rely “primarily or solely” on the output. Anthropic’s Claude for Healthcare (terms) similarly disclaims accuracy and places verification responsibility entirely on the user. FDA-cleared imaging AI vendors and ambient documentation tools like Nuance DAX and Abridge follow the same pattern.

An example of how this plays out in practice: in Sampson v. HeartWise Health Systems (386 So.3d 411, Ala. 2023), physicians followed cardiac screening software that misclassified a young adult with a family history of congenital heart defects as normal. The patient died. The court dismissed the developer’s liability because the licensing agreement gave physicians final decision-making authority — but it reinstated the negligence claims against the physicians themselves. The vendor escaped; the doctors did not. That is what output-related exclusions do in practice.

Vendors also draw the boundary between “decision-support” and “medical device” very deliberately to avoid higher-tier regulatory scrutiny. OpenAI’s Healthcare terms explicitly prohibit using the tool for real-time analysis of ECG waveforms, genomic sequences, or in vitro diagnostic signals — uses that would trigger FDA medical device classification. This boundary, between “decision support” and “medical device,” is partly a clinical distinction but it is also a legal one: staying on the decision-support side keeps the vendor out of device regulation and, not coincidentally, keeps liability with the physician.

The Confidence Problem

The liability model places full responsibility on the physician to catch AI errors before they harm patients. Whether that is reasonable depends heavily on how detectable those errors are. With a reference book, a physician has confidence cues: a publication date signals currency, a known author signals authority. With a lab test, the false-positive rate is documented and known.

Current AI systems, particularly large language models, strip away most of those cues. They present all outputs with similar fluency and authority regardless of whether the underlying confidence is high, moderate, or essentially nil. A well-grounded answer and a hallucination can be indistinguishable on the screen. The same query may produce different responses on different days. There is no reference list, no publication date, no documented error rate for a given clinical context.

This matters because the liability model assumes the physician can exercise meaningful oversight. When a tool’s failures are invisible — when it sounds equally authoritative whether it knows the answer or doesn’t — that assumption becomes harder to sustain. The physician is being asked to catch errors they have no reliable signal are occurring.

When There Is No Physician in the Loop

The physician-tool model described above presupposes a physician. When AI makes clinical decisions autonomously — without a physician reviewing the output — that model simply does not apply. There is no physician to bear responsibility. The liability question has to be answered some other way.

This is not hypothetical. There are four FDA-cleared autonomous AI diagnostic systems, all for diabetic retinopathy screening: LumineticsCore (Digital Diagnostics), EyeArt (Eyenuk), AEYE-DS (AEYE Health), and Retina-AI Galaxy. And in December 2025, Utah launched the first US program allowing AI to autonomously renew prescriptions without a physician involved in each transaction — developed by Doctronic, covering roughly 190 chronic-condition medications for patients who already have an existing prescription from a human clinician.

Digital Diagnostics and Doctronic have both answered the liability question directly: they accept it. Digital Diagnostics carries medical malpractice insurance for LumineticsCore and assumes liability for injuries arising from misdiagnosis. Doctronic secured a custom malpractice policy through Beazley, a global specialty insurer, explicitly covering the AI’s clinical decisions and holding the system to physician-level accountability standards. Utah’s director of AI policy was plain about the logic: “Generally, if it’s Doctronic’s product, it’s Doctronic’s liability.”

The other FDA-cleared autonomous diagnostic systems have not made their liability terms publicly available. Eyenuk’s website has standard terms of service covering website use; there are no public product terms, no statements about malpractice insurance, no disclosure of who bears liability when EyeArt misdiagnoses a patient. AEYE Health’s website has no visible terms of service at all. These are systems making autonomous diagnostic decisions without physician oversight, and the question of who bears responsibility when they are wrong cannot be answered from publicly available information.

What Buyers Should Ask

There is almost no AI-specific case law yet, and the US has no federal statute on AI medical liability. The insurance market is already moving: some carriers are introducing policy riders for heavy AI use, and European insurers are applying sublimits and higher deductibles for unvalidated tools. Health systems operating in this environment should be asking specific questions before procurement.

For AI used alongside physicians:

• What do the output-related exclusions actually cover? Understand that standard contracts place liability with the physician even when the AI’s output is the proximate cause of an error.

• What documentation does the vendor require or recommend? Given that courts will ask whether the physician’s decision to follow or reject AI output was reasonable, documentation of that reasoning matters.

• What training is in place to help clinicians recognize AI’s confidence problem — the fact that the output sounds equally authoritative whether the AI is certain or hallucinating?

• What audit trail does the AI generate? Documentation of whether and how the physician reviewed AI output will matter in litigation.

For autonomous AI without a physician in the loop:

• Who bears liability when the system is wrong?

• Does the vendor carry malpractice insurance for the AI’s clinical decisions? Digital Diagnostics and Doctronic do. If a vendor cannot answer that question, that is itself informative.

• Is the “autonomous vs. decision support” classification based on clinical design or legal positioning?

The Model Worth Watching

Digital Diagnostics has argued publicly that assuming liability is not just an ethical position but a business one: autonomous AI that accepts responsibility for its decisions gets reimbursed at a sustainable rate because it creates genuine, accountable value. CMS validated that argument by establishing a national reimbursement rate specifically for autonomous AI. Doctronic is testing the same thesis in prescription renewal, with a named insurer and a state regulatory framework built around the principle that the product’s creator bears responsibility for what it does.

The rest of the industry has mostly taken the opposite approach: disclaim the output, classify as decision support, and leave the physician holding liability for errors they may have had no reliable way to detect. That approach is legally coherent and, as Sampson shows, courts have upheld it. But it is a significant departure from the model that has governed liability across clinical tools for decades — one that physicians and health system buyers deserve to understand before they sign.

OpenAI for Healthcare

OpenAI's offering has two parts: ChatGPT for Healthcare (an end-user product) and the OpenAI API (for developers).

ChatGPT for Healthcare is a dedicated workspace where clinicians, administrators, and researchers log in, ask questions, and get cited answers from medical literature. It connects to Microsoft SharePoint, Outlook, and Teams, so responses can reflect an organization's approved policies and care pathways. Features include:

• Clinical search with citations from peer-reviewed studies and clinical guidelines

• Microsoft integration for organizational documents

• Reusable templates for documentation, prior auth, and patient communications

• Role-based access controls

The early adopters are large health systems with mature Microsoft environments: HCA Healthcare, Cedars-Sinai, Boston Children's, Memorial Sloan Kettering, AdventHealth, UCSF.

The OpenAI API lets developers build custom applications on top of GPT models with HIPAA-compliant infrastructure. Companies like Abridge, Ambience, and EliseAI already use it to power ambient documentation, scheduling, and care coordination tools. OpenAI also acquired Torch Health in January 2026, a startup that aggregates patient data from multiple EHR systems using FHIR APIs — suggesting the API platform will eventually offer its own connectors to clinical data sources.

Claude for Healthcare

Anthropic's offering also has two parts: Claude itself (an end-user product for Team and Enterprise customers) and a set of healthcare-specific connectors and developer tools.

Claude Team and Enterprise already connects to Microsoft 365 (SharePoint, OneDrive, Outlook, and Teams), so clinicians, administrators, and researchers can query organizational documents directly. There's no healthcare-specific branding or clinical search templates — it's the same Claude interface used across industries, with healthcare connectors layered on top.

The healthcare connectors are available to any paid Claude user — not just enterprise customers. A clinician with a Claude Pro subscription, for example, can enable them directly from the interface (see image at right).

As Anthropic's Eric Kauderer-Abrams explained at the launch: "These are not new models or even new products. What we're announcing is a coherent set of connectors so that Claude — our models and Claude our products, however you're accessing Claude — can work with healthcare data."

The healthcare developer tools use Anthropic's Model Context Protocol (MCP) to connect to clinical databases, plus "Agent Skills" — templates for common workflows. The connectors include:

• CMS Coverage Database (Local and National Coverage Determinations for Medicare)

• ICD-10 (diagnosis and procedure codes)

• National Provider Identifier Registry (provider verification and credentialing)

• PubMed (biomedical literature)

For life sciences customers, there are additional connectors to Medidata, ClinicalTrials.gov, bioRxiv/medRxiv, Open Targets, and ChEMBL.

The primary use case Anthropic emphasizes is healthcare IT teams using these connectors to build custom applications — a prior authorization tool, a claims appeal workflow, a care coordination system. Anthropic's partners reflect this approach: Banner Health, Elation Health, and Carta Healthcare (and many others) are building custom implementations on top of Claude — not using out-of-the-box products.

Competitive Advantages

Healthcare-specific connectors. Anthropic explicitly advertises connectors to CMS Coverage Database, ICD-10, and the NPI Registry. OpenAI's documentation doesn't mention these. It's very likely that some version of those databases is in ChatGPT's training data — so that, for example, ChatGPT for Healthcare knows the ICD-10 codes — but that's different from pulling live data from CMS, or the NPI registry, or other sources. For revenue cycle and prior auth workflows that depend on current coverage rules, this could matter. Or OpenAI may simply not have emphasized these features in their launch documentation. Worth clarifying before choosing a platform.

Healthcare packaging vs. general infrastructure. Both platforms connect to Microsoft 365. The difference is what sits on top. OpenAI bundles its Microsoft integration into a healthcare-branded product with clinical search, documentation templates, and role-based access controls — purpose-built for clinicians who want to log in and start working. Anthropic offers the same Microsoft connectivity as general enterprise infrastructure, then layers healthcare-specific database connectors on top for teams that want to build custom workflows. One is a finished product; the other is a toolkit.

Development capacity. Can health systems actually build effective applications using these tools? Most health systems aren't software companies. This limits adoption of both platforms' developer APIs, but OpenAI hedges this risk with ChatGPT for Healthcare — a healthcare-specific product that requires no development work to deploy.

Life sciences crossover. Anthropic's clinical trial and regulatory connectors give it a foothold in pharma and biotech. If those relationships pull Anthropic into health system partnerships through clinical trial sites, that's a vector OpenAI doesn't have.

What It Means for Health System Leaders

If you want clinicians using AI next month with minimal IT involvement, OpenAI for Healthcare is the more turnkey option. It's a healthcare-branded product with built-in templates and clinical search — connect your SharePoint and go.

If your organization already uses Claude Team or Enterprise, your staff can already query organizational documents through Microsoft 365. The healthcare connectors add clinical data sources on top of that. And if you have development capacity for custom workflows — prior auth, claims appeals, care coordination — Claude gives your IT team more healthcare-specific connectors to build with.

It seems inevitable that both companies will eventually offer the full stack. The question is what you need now.

A note on sourcing: Everything in this essay is derived from public documentation and launch announcements. I haven't used either platform in an enterprise healthcare setting. If you have direct experience deploying ChatGPT for Healthcare or Claude for Healthcare, I'd welcome corrections, additions, or context in the comments.

Part of what had prompted that exchange was that two significant AI health stories had all hit in the first week of January 2026 — new tools, each taking a very different liability approach.

1. Utah became the first state to let an AI autonomously renew prescriptions.

2. OpenAI launched ChatGPT Health, connecting the chatbot to your medical records and Apple Health data.

I've been writing about the migration of healthcare tasks from exam rooms to living rooms for a while now. What's new isn't that AI is doing health-related work. What's new is that the legal and regulatory frameworks are now being actively constructed, and the approaches are strikingly different.

Let me walk through what I found.

Doctronic: the cutting edge

Let’s start with the one that went furthest.

In January, Utah partnered with Doctronic to launch the first AI system in the country legally authorized to make clinical decisions about prescription renewals — 192 commonly prescribed drugs, $4 per refill. To make this possible, Utah's Office of Artificial Intelligence Policy signed a Regulatory Mitigation Agreement (RMA) with Doctronic in October 2025, under which the Division of Professional Licensing agreed to forgo enforcement actions for what would otherwise require medical licensure — specifically for this pilot, for 12 months.

This is the most forthright regulatory approach of the two. The state openly acknowledges that the AI is performing a clinical function. Doctronic has secured malpractice insurance holding its AI to the same standard of care as a human physician. And the RMA explicitly preserves patients' legal remedies: Section 4.D states that the agreement "does not waive or modify any legal remedies available to any individual harmed by any action of Participant's AI technology."

Doctronic also operates a separate consumer product — a health information chatbot available in all 50 states — with its own Terms of Service that take a very different posture: "Doctronic is an AI doctor, not a licensed doctor, does not practice medicine." The Utah pilot operates under the RMA, on an entirely different legal footing.

Basically, Doctronic evolved from a chatbot disclaiming medical practice to a state-authorized AI prescriber in a matter of months. But business moves quickly; as Zach Boyd, director of Utah's Office of AI Policy, put it: "We're in the Department of Commerce ... which is the entity that draws and redraws those lines all the time."

OpenAI: the disclaimer approach

OpenAI takes a very different tack. To understand the legal picture, it helps to know that OpenAI has two consumer-facing health-related products — ChatGPT and ChatGPT Health — and to understand exactly where the legal documents overlap and where they don't.

Start with what's the same. OpenAI uses a unified policy structure, meaning the safety standards and liability terms for a standard ChatGPT conversation and the dedicated ChatGPT Health feature (which is a separate tab in the interface) are governed by the same core documents. The Usage Policies are identical: you cannot use either one for "tailored medical advice" or "high-stakes" clinical decisions without a human professional. The Terms of Use are identical: both carry the same disclaimer ("not a medical device," "not for diagnosis," "informational only"), the same $100 liability cap, and the same assumption of risk. If you follow a suggestion from the Health tab that leads to a health issue, OpenAI's legal protection is exactly the same as if it happened in a regular chat.

The only document that is any different is the Health Privacy Notice, which applies exclusively to the Health feature. It promises that health data is segmented from your general chat history, sets a strict default that health content is never used to train foundational models, and governs the legalities of syncing with third-party sources like Apple Health and b.well (which aggregates data from 2.2 million U.S. providers).

In other words: you aren't "allowed" to do more medical things in the Health tab than in regular ChatGPT. The Health feature provides better tools (medical record syncing, lab result uploads) and stricter privacy (no model training) for the things you were already allowed to do. But the liability framework is unchanged.

This matters because real harm has already been documented from regular ChatGPT — not the new Health feature. A man was hospitalized for three weeks with bromism after ChatGPT recommended sodium bromide as a table salt substitute. Warren Tierney, a 37-year-old Irish father, delayed cancer care after ChatGPT repeatedly assured him it was "highly unlikely." Neither of these has resulted in litigation, but separately, seven wrongful death lawsuits were filed against OpenAI in November 2025, including Raine v. OpenAI, alleging that ChatGPT encouraged a teenager's suicide. Those cases are testing whether a $100 liability cap and a disclaimer can hold when the product increasingly functions as a health advisor. And now the product can ingest your full medical record — under the same legal terms.

It's also worth noting that in October 2025, OpenAI updated its usage policies to prohibit "provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional." Then two months later, they launched ChatGPT Health — a product specifically designed to provide personalized health information by ingesting your medical records. Make it make sense.

[Note: ChatGPT for Healthcare is a separate enterprise product for health systems governed by OpenAI's business Services Agreement with a Business Associate Agreement for HIPAA compliance. Different product, different market — hospitals and clinics, not consumers. More on this coming up.]

Comparing two consumer-facing product approaches

The bigger picture: no size fits all

What you're looking at is an evolving field where consumer-facing AI health tools are emerging faster than any single regulatory framework can accommodate — and different actors are responding in very different ways. Doctronic sought state authorization and malpractice insurance. OpenAI added clinical capabilities without changing their liability terms.

And the landscape is expanding. Doctronic is already in talks with Arizona and Texas. H.R. 238 (introduced in January 2025) would create a federal framework recognizing AI as a "practitioner licensed by law to administer drugs." The wrongful death cases against OpenAI are testing the disclaimer approach in court.

What's notable is how differently these two companies are responding — Utah's proactive sandbox versus OpenAI's unchanged liability terms.

So who's liable when consumer-facing health AI makes a mistake? It depends entirely on which product you're using. If Doctronic's AI renews the wrong prescription in Utah, there's malpractice insurance and the RMA preserves your right to sue. If ChatGPT Health steers you wrong after ingesting your full medical record, your recourse is capped at $100 — the same as if you'd asked it about a recipe (if that holds up in court). The answer to the title question isn't "nobody" across the board. It's that the answer varies enormously depending on which product — and that's the landscape we're in right now.

This post draws on themes from my keynote talks.

→ Follow new essays via RSS or on LinkedIn.

AI models like ChatGPT, Claude, and Gemini are getting serious attention for what they can do that improves on the traditional healthcare system. As the New York Times recently noted, while imperfect, they are "empathetic, available, cheap." They can explain complex medical concepts, analyze symptoms, interpret lab results, and provide personalized health guidance that would have seemed like science fiction five years ago.

But these AI models share a critical flaw with the traditional healthcare system: they're episodic. They only work when you're actively engaging with them. They forget context between sessions, can't run tasks in the background, and never take initiative. The healthcare system does have a proactive aspect - annual checkups - but once a year is painfully far from real-time monitoring. Between those visits, both AI chatbots and traditional care are reactive, waiting for you to notice something is wrong.

For management of our health over time, this episodic model is poorly fit for purpose. And that's true whether you have a chronic condition or you're generally healthy and want to stay that way.

This month, I've been watching OpenClaw, a not-ready-for-prime-time but hugely popular nerd project that nonetheless illustrates how we might solve these problems. (For background on OpenClaw's rapid rise and security concerns, see TechCrunch's overview.)

What OpenClaw Does Differently

OpenClaw is a personal AI that runs continuously on your own computer. Unlike chatbots, it runs in the background, remembers everything, performs scheduled tasks automatically, and takes initiative without being asked. For healthcare, where continuous monitoring matters more than episodic conversations, this architecture changes everything.

1. Automatically gathers data

Instead of waiting for you to upload data, OpenClaw can check your glucose monitor every 15 minutes, pull activity data throughout the day, monitor your pharmacy for refills, watch for lab results in email, and prepare summaries before doctor appointments. Your data gets collected and acted upon continuously without you remembering to track anything.

2. Cross-references data

Apple Health and Fitbit show charts and send basic alerts - heart rate went up, sleep score declined. But they don't connect dots across multiple data sources. OpenClaw can cross-reference everything: your step count drops 40% on days when sleep quality is poor. It could even connect grocery purchases (via store loyalty apps) to activity, sleep quality, or diabetes risk — which as far as I know has never been done at any scale.

3. Communicates intelligently and acts on your behalf

It communicates several ways:

• Passive dashboards when you want to check in

• Alerts when patterns emerge: "Your resting heart rate has been elevated for three days - this could be an early sign of infection, or stress"

• Followup: "If you think either of those could be possible, we can explore next steps"

and maybe the best for last:

• Actions on your behalf: OpenClaw could automatically schedule follow-up appointments when lab values fall outside ranges, create pre-appointment summary brief to show to doctors, order prescription refills, create shopping lists based on dietary restrictions, and adjust smart home settings based on sleep data.

The Lobster Cometh

For more and more people, the present is something like "with substantial effort, I use wearables, portals, and AI to better understand my health data, track symptoms, and prepare for doctor visits."

With tools like OpenClaw, we can imagine a future more like "I have a continuous AI assistant handling data collection, analysis, specialist coordination, and scheduling. It only alerts me when there's an issue or decision needed."

OpenClaw is too insecure, and too technical for mainstream adoption today. But it's a preview of what's coming — and of demand: it went from obscurity to 2 million weekly users in weeks.

You can bet Apple will integrate this, or something like it (iClaw? Apple Claw?) into the Health app. Specialized health AI companies will build HIPAA-compliant versions for chronic conditions. Consumer AI platforms will add always-on agent modes.

It's not a question of whether this will happen: it's happening now.

Related reading:

• From Exam Room to Living Room: The New Health System, Part 1

Have you experimented with OpenClaw or similar agent frameworks? I'm particularly interested in healthcare use cases and security considerations. Reach out at joel@futurehealth.live.

This is a lesson that healthcare has not yet absorbed. They think that Epic, leading EHR maker, is a health software giant that will lead in applying AI to our health—not realizing that Apple and Google are now health companies and orders of magnitude larger. As I wrote in Forget the EHR — Your Health Data's On Your Phone:

What does it mean for the future of health when Apple makes more money from AirPods than the top ten EHR vendors combined?

That comparison matters because consumer technology companies have become health companies—whether they call themselves that or not—with more money, more programming talent, and more access to the population than any EHR company or healthcare system.

This is even more significant in the AI era, because AI innovation requires data, and we now have 50,000 times more consumer health data than EHR data, simply because there are more consumer computing devices—including phones constantly measuring and recording our speed, direction, and frequency of movement.

And this has caused a decades-old, large-scale migration of health-related activity from the healthcare system to the consumer tech system—to devices, apps, and AI with no provider involved at all.

Five Tasks Migrating Out of Healthcare

One way to think about the migration is by core healthcare tasks: knowledge, diagnosis, treatment, monitoring, and interpretation. Over decades, portions of each have steadily migrated outward from clinics into homes. Not all at the same speed, but all moving in the same direction—from clinic to home. From provider-mediated to no-provider-involved.

And never the other way around.

Knowledge

Patients have always been interpreters of their own symptoms. What changed with the Web was the amount of context available before involving a clinician. Suddenly, around 25 years ago, consumers had access to the same medical reference material as clinicians, via websites like drugs.com, Medscape, and of course Google.

None of this replaced clinicians—it simply gave the patient information. But across multiple studies in systems like Kaiser, the NHS, and others, a universal finding is that giving consumers reliable medical information consistently reduces low-acuity in-person visits—millions per year in the US.

Diagnosis

It started in the 1970s with home blood pressure cuffs, primitive glucose meters, and early pregnancy tests. Now the category has exploded: direct-to-consumer lab testing, rapid covid and flu tests, continuous glucose monitors. Twenty million home pregnancy tests are sold annually in the US—each one representing a service that once required a clinic visit.

Consumer devices now perform complete diagnostic functions that once required clinic visits and specialized equipment. The Apple Watch can detect atrial fibrillation and, with new apps, perform a complete sleep study—not just collect data for a physician to interpret, but deliver the diagnosis to the user. Hearing tests used to require trained personnel and expensive equipment; now you can do them with AirPods and your iPhone.

Consumer tech is also providing health services that were never really practical before: Samsung, for example, has announced that its phones will be able to screen for early signs of dementia from the language we use in our communications. Every day, all day, not just at a once-a-year checkup.

Treatment

Just this month (January 2026), Utah became the first state to allow AI to legally write prescription renewals for chronic medications: Doctronic’s AI platform, with no human provider in the loop, can now renew prescriptions autonomously within the state's regulatory sandbox. Arizona, Texas, and Wyoming are exploring similar approaches. Back in 2019, the FDA under Scott Gottlieb was reported to have been looking at this kind of "autonomous prescribing," and those discussions are becoming reality.

But AI-assisted treatment is just the latest chapter in a much longer story that predates AI. In Disruption for Doctors 3: The Rise of Selfcare, I calculated the scale of this shift:

More than 700 medications now available over-the-counter in the US once required a doctor's prescription. Zantac, Flonase, Claritin, Advil—each was once a billable clinical decision. A 2012 Booz [now PwC] study estimated that without OTC medications, we'd need roughly 60,000 additional clinicians just to staff the extra visits.

People often ask me whether AI will replace doctors—but no one ever asks whether OTC drugs will replace doctors. Even though that is precisely what they're designed to do, and precisely what they're doing.

Current FDA Commissioner Marty Makary has been vocal about moving more medications to OTC status. He clearly understands what most healthcare observers miss: OTC conversion isn't a minor consumer convenience. It's a systematic transfer of treatment decisions out of the clinical encounter.

Of course, treatment is not just pharmaceutical treatment. Consumer tech is also moving into physical therapies: the non-prescription Osteoboost Belt, cleared in 2026, delivers vibration-based bone stimulation to slow progression of osteopenia—a condition driving $19 billion in annual U.S. healthcare costs. And it seems clear at this point that consumer tech will be able to provide affordable and effective treatment for a variety of mental health conditions as well.

Monitoring

For most of the 20th century, monitoring meant having something measured during a clinic visit—blood pressure, weight, maybe a lab—once or twice a year. Once sensors moved onto wrists and into pockets, monitoring became continuous. Sleep patterns, heart rhythm, glucose trends, movement patterns, fertility windows—all visible in real time, all day, every day.

From Digital Coaches Part II: Prevention's New Business Model:

Activity trackers like Apple Watch, Fitbit, and Garmin now deliver personalized, continuous feedback at global scale, turning prevention into a business that keeps people healthier—and needing healthcare less.

Continuous monitoring has transformed how well health decisions can be made. Trends replace guesswork. Early changes become visible. People adjust long before a clinician would ever see the issue.

Interpretation

Interpretation addresses questions that traditionally required a clinician's training: synthesizing signals; evaluating trends; relating observations to the patient's age, history, medications, and risk factors; and comparing all that with the hundreds or thousands of patients previously seen.

One simple example is Apple's Walking Steadiness feature—one of the reasons I bought my 85-year-old mom an iPhone. Introduced in 2021, this wasn't launched as a medical breakthrough or promoted at all. It simply appeared one day on a device already used by millions of older adults, at no additional cost. Yet consider what it does: silently monitors movement patterns, detects concerning trends, explains what those patterns typically mean, suggests exercises that might help, tracks whether things improve, and can notify family members if they don't.

On the more complex end, from The Rise of Digital Health Coaches:

AI health coaches aren't just for athletes anymore. They're starting to handle the day-to-day interpretation, advice, and treatment adjustments that once required doctors. From glucose monitoring to hypertension management, technologies like Dexcom, Teladoc, and Omada are quietly taking over the work of routine clinical decision-making.

Just in the past few months, several AI providers have launched health functionality that can look at ALL of your health data, including from your watch, your glucose monitor, your patient portal, your smart scale, and more. These products, such as ChatGPT Health, can then help you understand test results, prepare for appointments, get advice on diet and exercise, or understand insurance tradeoffs based on your healthcare patterns.

When Will Healthcare Notice?

Outside healthcare, consumers are assembling their own systems: wearables, apps, AI tools, pharmacies, online communities, and self-directed care pathways. None of these pieces looks like a hospital. But together, they perform an increasing share of the work that the hospital used to do.

This system is fragmented, uneven, and imperfect. But it's improving faster than institutional healthcare IT ever has, because it rides that all-powerful consumer technology wave that transformed every other sector.

Importantly, consumers aren't using those tools because they necessarily think they're better than seeing a human doctor. They're using them because of convenience, cost, speed, and control. For many everyday health-related jobs, those factors matter more than perfection.

You might think that the remarkable advances in AI in the last few years would have pushed healthcare to finally recognize these migrations and the challenge they pose. But in my experience speaking with hundreds of healthcare CEOs and board members over the past five years, the opposite is true: these migrations remain largely invisible to healthcare leadership. The industry is focused on competitors it can see—retail clinics, DTC websites, telehealth startups. It's missing the much larger shifts happening without any provider involved at all.

Why can't healthcare see what's happening? That's the subject of Part 2.

Coming soon: The Competition Healthcare Can't See, The New Health System, Part 2

The 2019 version of this guidance covered wellness claims but said nothing about devices that measure clinical parameters. The assumption at the time was simple: if you measure blood pressure or glucose, you're probably a medical device.

The 2026 update says: not if you're coaching for wellness. Measure anything you want non-invasively, provide detailed feedback, offer AI-driven recommendations—just don't claim to diagnose or treat disease.

In October, I wrote about digital health coaches—AI systems handling day-to-day interpretation, advice, and treatment adjustments that once required doctors. In November, I explored how that coaching is giving prevention a business model at global scale through devices like Apple Watch and Oura Ring.

Well, the FDA just gave those products an easier regulatory path for further innovation.

This matters because the wellness device market is enormous. Apple sold over 50 million Apple Watches in 2024. Dexcom has 1.6 million users. WHOOP, Oura, Levels, Ultrahuman—millions more. These companies can now measure clinical parameters and provide coaching without navigating medical device regulations.

One less reason to visit the regulator.

What Changed

The updated guidance expands what counts as "general wellness." It now explicitly covers non-invasive sensors that estimate physiologic parameters—as long as they're intended for wellness use only.

These devices can:

• Display values, ranges, trends, or longitudinal summaries

• Contextualize outputs in relation to sleep, activity, stress, recovery

• Notify users when outputs fall outside ranges appropriate for wellness

They can't:

• Claim clinical equivalence or medical/clinical grade accuracy

• Provide alerts requiring specific clinical action

• Offer treatment guidance for medical management

• Reference diagnostic thresholds or specific diseases

A wearable that monitors pulse rate, blood pressure, sleep quality, and activity for "assessing recovery" qualifies as wellness—as long as it doesn't claim to diagnose hypertension or replace a medical device.

The Contradiction

There's an inherent tension in the new guidance: it says you can't claim to be treating disease. The problem is that although lots of "wellness" is marketing fluff, the core — exercise, sleep, and eating well — in fact IS treating disease.

For example, exercise is first line therapy for hypertension. So if a device measures your blood pressure, coaches you to exercise because it's trending upwards, and your pressure drops as a result . . . isn't that treating hypertension?

The FDA's answer: not if you don't claim it is.

You can say exercise "may help manage high blood pressure" but not "treats high blood pressure." The biological effect is identical. Only the framing differs.

This isn't limited to hypertension. Sleep coaching can lower diabetes risk—but don't call it diabetes prevention. Stress reduction improves cardiovascular outcomes—but don't call it cardiac therapy. Weight loss from activity tracking reduces joint pain—but don't call it osteoarthritis treatment.

Core wellness interventions certainly have medical effects, but the FDA is willing to pretend they don’t, in order to allow more innovation.

This works for now. It won't work forever. Consumer wellness tech advances faster than medical tech because more money and talent is going after it. As wellness devices get more sophisticated, it's likely that the treatment they offer will become more effective, and the gap between (for example) "wellness coaching that improves your blood pressure" and "treating hypertension" will be even smaller.

Utah and FDA: Two Parts of the Same Regulatory Story

The same week FDA released their wellness guidance, Utah became the first state to let AI renew prescriptions with no doctor involved.

Both are regulatory relaxation. Utah created a sandbox—temporarily relaxing laws to let companies practice medicine with AI. The FDA clarified that wellness devices measuring clinical parameters don't need permission at all.

Different mechanisms, same result: moving healthcare tasks out of traditional healthcare systems.

Utah's sandbox doesn't just apply to hospitals or traditional healthcare companies. It applies to DTC companies like Ro and pharmaceutical companies' direct-to-consumer efforts like Lilly Direct. These aren't healthcare systems—they're consumer companies that happen to deal with health.

By reducing doctor workload (no more refills to deal with), the UT rule means that fewer doctors will be required per unit of work. And who will be able to respond more quickly to that reduced requirement: traditional healthcare? Or DTC startups like Ro or DTC pharma like LillyDirect? Who will more quickly, and readily, reduce their physician staffing levels?

Obviously, the DTC companies will move faster. They're not built around physician labor. They're built around technology that replaces physician labor.

Traditional healthcare systems will debate AI's role in clinical workflows while consumer companies eliminate the workflows entirely.

Both the FDA guidance and Utah's sandbox accelerate the same trend: health moving to consumer channels where traditional healthcare can't follow.

The Bigger Picture

Health-relevant capabilities keep migrating from clinic to consumer. OTC drugs, home tests, smartphone diagnostics, now AI-powered wellness coaches with physiologic monitoring.

Each migration follows the same logic: if the task doesn't require medical judgment for safety (or no longer requires it), let tech-enabled consumers handle it themselves. And FDA's updated guidance plus Utah’s relaxed regulation just expanded the category of tasks that don't require medical judgment.

Healthcare systems worry about retail clinics and urgent care centers as competitors. The real competition is coming from devices that keep people out of healthcare entirely. Not by denying care, but by providing it elsewhere.

Utah's sandbox and FDA's guidance are both accelerating that shift. Both increase the number of things that don't require a human expert -- or a traditional healthcare system -- to provide. Both move healthcare tasks to consumer channels. Both reduce the need for traditional healthcare infrastructure.

The market is already moving. Utah and FDA just cleared the path a bit more.

This is the third piece in a series on digital health coaches, wearables, and consumerization of health. See also The Rise of Digital Health Coaches and Digital Coaches, Part II: Prevention's New Business Model.

Blue Cross Blue Shield of Massachusetts ended 2024 with a $400 million operating loss — the worst in the organization's history. The primary driver wasn't hospital costs, physician fees, or even the lingering effects of the pandemic.

It was GLP-1 medications.

The drugs — Ozempic, Wegovy, Mounjaro, Zepbound — accounted for more than $300 million in spending, representing 20% of the insurer's total pharmacy costs and double what they spent the year before.

Massachusetts isn't alone. North Carolina's state employee health plan canceled GLP-1 coverage for weight loss entirely — saying that continuing it would have required doubling premiums for all 750,000 members. Colorado stopped covering new GLP-1 prescriptions, grandfathered existing users, and quadrupled their copays.

Here's the paradox: these same drugs are already delivering population-level results.

The Evidence Is In

This isn't theoretical anymore. According to Gallup's National Health and Well-Being Index, U.S. adult obesity has dropped from 39.9% in 2022 to 37% in 2025 — representing approximately 7.6 million fewer obese adults. The decline seems to correlate directly with GLP-1 adoption: the share of adults using these medications for weight loss has more than doubled, from 5.8% in early 2024 to 12.4% today.

And GLP-1 receptor agonists don't just help people lose weight. They appear to reduce cardiovascular events, protect kidney function, and — emerging data suggests — may even slow neurodegenerative disease (or maybe not).

We’re already seeing fewer bariatric surgeries. If these medications continue working at population scale, we should see fewer diabetes complications, fewer heart attacks, fewer joint replacements in patients whose weight made them poor surgical candidates. The downstream savings could be enormous.

But "downstream" is the operative word.

The Timing Problem

The pharmacy costs hit immediately. The medical cost savings — if they materialize for any given insurer — arrive years later. And in a health insurance system where members frequently change plans, the insurer paying for the drug today may not be the one who benefits when that member avoids a heart attack in 2031.

This is bad timing, not bad economics: a therapy that is eliminating demand for expensive downstream care is financially crushing the organizations that would presumably benefit from that demand reduction.

One counterargument: With 42% of American adults obese, every insurer has similar exposure to these costs. They're all bleeding together — and in aggregate, should all benefit together as the obesity rate continues to fall. But that's cold comfort when you're explaining a $400 million loss to your board.

What This Means for Health Systems

I've written before about how GLP-1s represent a potential demand shock for health systems — the first large-scale collapse in chronic disease volume in living memory. But if you're a health system executive, you might be tempted to see this insurance crisis as someone else's problem: let the payers figure out how to pay for it.

But there are serious implications for folks on the care side of the equation:

The insurers' eventual savings are your volume losses. When GLP-1s reduce cardiovascular events and diabetes complications, that's fewer procedures, fewer admissions, fewer chronic disease management visits flowing through your system. The math that eventually works for payers works against traditional fee-for-service health systems.

Coverage is contracting, not expanding. As losses mount, payers are pulling back. Employer coverage for GLP-1s for obesity faces pressure. Many insurers are adding prior authorization requirements or dropping coverage entirely. Blue Cross MA just announced it will stop covering GLP-1s for weight loss effective January 2026. The brief window of broad access may be closing.

The DTC market is filling the gap. Millions of Americans have obtained compounded GLP-1s through direct-to-consumer telehealth platforms — often after nothing more than a questionnaire. No lab work. No ongoing monitoring. No care coordination. The demand hasn't disappeared; it's migrating outside the system entirely.

Demand elimination is coming regardless. Whether patients get GLP-1s through your system, through their insurer, through Amazon, or through a compounding pharmacy in Florida, the downstream effects on surgical volumes and chronic disease management will arrive. The question isn't whether demand elimination happens — it's how you are planning to respond to it.

The Strategic Question

For health system boards, this moment demands clarity about a fundamental question: Are you primarily in the business of treating disease, or managing health?

If you're in the disease treatment business, GLP-1s represent a threat. Fewer obese patients means fewer bariatric procedures, fewer obesity-related joint replacements, fewer cardiovascular interventions. Your volumes decline.

If you're in the health management business, GLP-1s represent an opportunity. You become the trusted partner helping patients access, monitor, and optimize these therapies — capturing the value of the intervention rather than waiting for the complications.

The insurers bleeding cash right now made a bet: cover these medications, absorb the short-term losses, reap the long-term savings. That bet is failing because the economics don't work on a single-payer timeline and because patients don't stay with the same insurer for a decade.

Health systems have a different hand to play. You have continuity. You have the clinical relationships. You have the ability to integrate medication management with the downstream specialty care that may (or may not) still be needed.

The question is whether you're building that integration now, while you still have the patient relationship — or waiting until they've gotten their prescription from a telehealth app and only show up when something goes wrong.

The Shift Is Here

Blue Cross Blue Shield of Massachusetts didn't post a $400 million loss because of a temporary blip. They're experiencing the front edge of a structural, not cyclical, shift in how obesity — and eventually other chronic diseases — gets treated.

The bad timing won't resolve itself quickly. Years of costs will add up before the savings arrive. The payers will continue to struggle with the math. And the patients will continue to find ways to access these medications, with or without the traditional healthcare system's involvement.

—

Sources

Blue Cross Blue Shield of Massachusetts $400M loss

Gallup obesity data (39.9% → 37%)

North Carolina canceled coverage

Colorado coverage changes

When the polio ward disappeared

““We have succeeded in eradicating smallpox from the world, but in doing so we have put a lot of people out of work—vaccine makers, field workers, and public health officials who dedicated their lives to this fight.” — D.A. Henderson, Director, WHO Smallpox Eradication Program

Every so often, healthcare experiences a period when demand doesn’t just shift or grow — it collapses. Clean water systems produced moments like this in the early 20th century, when filtration and sewage systems reduced the infectious diseases that once filled pediatric wards, slashing infant and child mortality. Vaccination did the same repeatedly across the 20th century, most visibly with the disappearance of polio wards by the late 1970s in the United States, and with the steep declines in measles, diphtheria, pertussis, and other once-routine childhood killers (though in recent years, declining vaccination uptake has sadly brought some of these risks back).

These were demand shocks of an extraordinary scale — entire categories of disease, and the care built around them, effectively vanished. From 1900 to 1980, annual infectious-disease deaths in the United States dropped by more than 95%. The effect on the healthcare system of the time was profound: diseases that had shaped the daily routine of hospitals — diarrheal illnesses, typhoid, diphtheria, measles, pertussis, tuberculosis — either shrank dramatically or nearly disappeared. Entire service lines lost their purpose.

Because these events unfolded decades ago, we rarely think of them as disruptions to the healthcare system. We remember them only as the public-health triumphs that they truly were, neglecting the healthcare-system-altering contractions they also represented.

GLP-1 drugs may represent our first chance to see something like this happen up close: triumph and challenge rolled into one.

From infectious disease to chronic dependence

That earlier collapse in infectious-disease burden happened while healthcare was still relatively small and still evolving. Hospitals, specialties, and payment structures were in their formative stages. As the disease landscape changed, the system grew in new directions: toward surgery, obstetrics, trauma, and, eventually, toward the chronic cardiometabolic conditions that dominate today.

What emerged over the second half of the 20th century was a healthcare industry whose volumes and revenues depend heavily on long-lasting, slowly progressive disease. Diabetes, obesity, hypertension, atherosclerosis, heart failure, sleep apnea, degenerative joint disease — these are no longer side stories. They are the central plot of modern healthcare.

That dependence is of course not just clinical, but financial. Service lines, buildings, staffing models, and capital plans all assume that cardiometabolic disease will be both prevalent and durable. In effect, the system has built an enormous, expensive apparatus around the expectation that these conditions will be with us, at roughly current levels, forever.

This is the backdrop against which GLP-1s arrive.

GLP-1s and the new demand shock

If GLP-1s significantly reducing obesity and diabetes, and everything downstream, that would touch nearly every major service line in today’s system:

• Internal medicine and primary care: fewer long-term hypertension visits, fewer chronic medication adjustments, fewer metabolic-syndrome check-ins, and smaller chronic-disease panels overall.

• Cardiology: fewer heart attacks and unstable-angina cases, fewer stent placements, fewer heart-failure admissions, and a smaller pool of high-risk patients.

• Endocrinology and diabetes care: fewer diabetes-management visits, fewer insulin starts, fewer specialist referrals for complications.

• Orthopedics: fewer knee and hip replacements driven by excess weight and joint stress.

• Bariatric surgery: smaller candidate pools and lower procedural volume.

• Sleep medicine: fewer sleep studies and CPAP initiations as apnea severity declines.

• Nephrology: fewer patients progressing to kidney failure through diabetic or hypertensive pathways.

• Wound care and vascular surgery: fewer amputations, fewer chronic wounds, fewer revascularization procedures.

In each of these domains, the issue is less disease arriving at the door, year after year. Spread across a decade or two, these reductions would reshape the economic foundation of many health systems.

What happens when your core customers gets healthier without you?

This is the strategic problem GLP-1s create for healthcare: the healthier your population becomes, the less they need the very services your system was built to provide. And if the disease burden shrinks — even gradually — the business model built on managing it must change. Not because care becomes more efficient, but because there is less care to deliver. And unlike the early-20th-century declines in infectious disease, today’s healthcare system doesn’t have the flexibility, margins, or structural slack to absorb a long, steady reduction in the conditions that sustain it.

The question facing health systems is therefore not, “How should we treat patients on GLP-1s?” but, “What happens when the chronic-disease core of our business begins to lighten?”

That is the adaptation problem most leaders have not yet begun to model.

How hospitals can adapt to less disease

So what does adaptation look like in a world where metabolic disease begins to recede?

First, boards and executives will need to take the possibility seriously enough to model it. I’ve spoken with more than 50 healthcare system CEOs about the GLP-1s meds and the possible financial effect on healthcare, and haven’t had a single one tell me they’re working on it. This means no one is asking “what if the GLP-1 meds do exactly what they seem to do?” What does your financial and clinical picture look like if, over ten to fifteen years, the prevalence and severity of obesity-related conditions in your catchment area decline meaningfully? Which parts of the system end up built for more patients than are showing up?

Second, systems will need to rebalance toward areas of enduring or rising need. Even in a healthier cardiometabolic world, there will be no shortage of dementia, cancer, complex geriatric care, behavioral health,, and palliative needs. Many of these are currently under-resourced relative to demand. If GLP-1s bring fewer diabetic amputations and fewer heart-failure admissions, that becomes the moment when health systems must decide whether the resources they free up simply evaporate — or get deliberately redirected.

Finally, this is an opportunity to align with demand elimination rather than resisting it. If GLP-1 meds continue to prove effective on-balance, and continue to improve, the smart move might be to integrate them: building programs around metabolic health, monitoring, side-effect management, and long-term risk reduction, even as the need for traditional downstream procedures shrinks. Though, as I’ve mentioned in other posts, healthcare systems that do this will be in competition with lots of startups trying to do this with wearables, AI, and other new tech.

The choice ahead

GLP-1s may trigger the first major decline in chronic-disease demand that modern healthcare has ever faced. The question is whether hospitals and health systems prepare for that shift by redesigning capacity, workforce, and strategy — or simply wait for volume to erode.

The systems that plan for a world with less metabolic disease will be the ones that remain viable in it.

The New York Times is often held up nowadays as one of the great success stories in an otherwise decimated journalism industry. Articles highlight growing subscriber numbers, rising digital revenue, and a product that looks far better than the paper of twenty-five years ago.

But if you look at a chart of their revenue and profit over that period, the pattern is unmistakable: real revenue is down nearly sixty percent. Real profit is down about seventy percent:

So while it’s true that the Times has built one of the best digital products in media, reached more readers than ever, and maintained a high standard of journalism — it’s also true that it is a shadow of its former self.

Disruption doesn’t kill the work — it moves it

In a nutshell: faced with digital disruption, the Times made a better product, delivered it to a bigger audience, and ended up with a lot less money.

That combination — better product, shrinking economics — is exactly what disruption looks like in the real world. The work that used to sustain the industry is now done somewhere else, by someone else, and — at least in the case of digital disruption — often for free.

Steel followed this pattern. Music followed it. Newspapers followed it. In every case, the consumer gained access and capability, while the incumbent industry saw its total market shrink. The integrated mills didn’t expand when minimills took the easy jobs. Record labels didn’t expand when digital distribution flattened the business.

And the internet didn’t erase the functions of a newspaper. It just pulled them out of the newspaper business. News now breaks on Twitter, analysis lives on Substack and podcasts, classifieds moved to Craigslist and Facebook Marketplace, and most daily information comes from apps. The work didn’t vanish. It simply moved outside the industry, leaving the industry itself much smaller.

Just history repeating

By all indications, healthcare is heading down the same well-worn path of disruption.

A large fraction of modern healthcare consists of informational tasks: explaining symptoms, interpreting labs, adjusting medications, and handling the low-level worries that humans generate every day. And as AI models, sensors, and apps continue to improve, a great deal of this informational work will move outside healthcare entirely. Not to another clinic or another clinician — simply out of the system.

Even if that were the only force at work, it would be enough to shrink the system. But consumer health technology isn’t the only powerful force at play here.

Winter is coming

It is becoming clear that GLP-1 medications and other metabolic therapies will reduce the number of people who ever develop the conditions that currently drive much of healthcare’s revenue: diabetes, heart failure, stroke, chronic kidney disease, severe obesity, osteoarthritis requiring surgery, several cancers, and the cascade of complications that accompany all of them.

In fact, Swiss reinsurance giant Swiss Re projects that GLP-1 drugs could by themselves cut US mortality by as much as 6.4% over the next twenty years (!).

Of course, the people who do reach the hospital will still need expert, highly specialized care — the kind that only major medical centers can provide. Cardiac surgery, transplants, trauma care, critical care, oncology, neurosurgery: none of that disappears. It remains essential.

But fewer people will arrive there if their health is better to begin with.

And it isn’t just GLP-1 meds.

Driver-assistance and autonomous driving systems are steadily reducing serious accidents — one of the biggest sources of high-acuity utilization. Fewer crashes means fewer surgeries and fewer long recoveries. Add this to consumer health disruption and the GLP-1 meds and the direction is obvious: people will need less from the healthcare system.

In a nutshell: consumer health tools are beginning to replace the everyday questions, new drugs are starting to reduce the chronic diseases that still dominate healthcare’s workload, and safer vehicles are steadily reducing the accidents that still generate so many trauma visits and surgeries.

Healthcare has no Wordle - and that’s a problem

When demand for newspapers collapsed, the Times cushioned the decline by expanding into things that had nothing to do with journalism: games, recipes, product reviews. Wordle didn’t save the economics of journalism, but it created a daily ritual that kept people connected to the brand even as the underlying business shrank.

Healthcare has nothing comparable.

No sideways expansion into a consumer-friendly category.

No Wordle or NYT Cooking or Wirecutter.

If you squint, you might imagine a hospital-branded, medically supervised fitness center, or nutrition programs built around metabolic science. Interesting ideas, but not likely to succeed at scale, and not playing to healthcare’s strengths.

As noted, healthcare does have a core that will remain — complex surgery, trauma care, critical care, oncology. But that core isn’t a cushion; it’s the remnant. And even that remnant will shrink as ADAS reduces trauma, GLP-1s cut metabolic disease, and AI and robotics continue reducing the need for invasive procedures. None of this high-acuity work can become healthcare’s Wordle. It isn’t lightweight, pleasant, cheap, or scalable. It can’t serve as a daily — or continuous — ritual that keeps millions connected to the system when they don’t need the system.

A Better Healthcare System — and a Smaller One

The New York Times today is a much better product and a much smaller company — and it’s the success story of journalism, an industry in which many newspapers simply disappeared, while others have just become a place to view Google ads.

Healthcare is now moving along the same trajectory, except the forces pushing it downward are stronger and the sideways options are fewer. More questions answered outside the system. Fewer chronic diseases progressing to advanced stages. Fewer car crashes sending patients to the trauma team.

We are heading toward a world where people truly need less from the healthcare system.

That is good news for our health, but it will not feel like good news to the institutions built for a very different era of demand. A system built on high volume, chronic disease, and constant incoming problems is not prepared for a future defined by less of all three.

The uncomfortable question — and one that healthcare leaders will eventually have to confront — is this:

What does a health system plan for when its business model shrinks faster than its mission?

That question shapes the future of healthcare.

And it is the lesson of the Times.

I ended that post by promising to explore how consumer-focused digital coaches are reshaping health for everyone—not just patients. These coaches aren’t directly challenging healthcare, but in the long term may reduce dependence on it.

Prevention has always been better — but harder

“An ounce of prevention is worth a pound of cure.” Everyone knows the phrase. Few systems live by it.

Healthcare has always been organized around illness — responding after something goes wrong. Prevention works earlier, acting before illness occurs.

Traditional public health achieves that prevention through collective infrastructure, regulation, and policy: clean water, sanitation, vaccination, pollution control. The new prevention is digital, built on wearables and other consumer technologies. It doesn’t replace those foundations — we still need vaccines and clean water — it extends them to the individual level.

It focuses on what public health has always found hardest to reach: the daily behaviors and decisions of individuals, shaped by personalized, continuous feedback and advice.

The new prevention sits outside both healthcare and public health, reaching people directly through their devices — every day. It’s the missing last mile of behavior.

From chronic disease to continuous prevention

The first generation of digital coaches grew up inside medicine. Companies like Livongo (now part of Teledoc), Omada, and Virta use connected sensors and human coaches to help diabetic or hypertensive patients stay on track between healthcare visits.

The next generation has moved beyond disease. It focuses on continuous prevention — helping people improve their activity, the behavior most reliably measured by today’s devices and most consistently linked with long-term health.

Inside healthcare, prevention was something doctors advised but rarely measured or were reimbursed for. Outside healthcare, it has become a major product: quantified, personalized, and available to everyone with a phone (which is to say, everyone).

This is both because the addressable market includes both the sick (as with healthcare) and the well, and because it doesn’t require dealing with healthcare’s substantial regulation and paperwork.

The new consumer coaches

A growing group of products now provide preventive guidance and feedback based on activity. Together, these products represent prevention delivered continuously, outside healthcare and without clinicians. Meta-analyses across brands confirm increases in daily steps and moderate-to-vigorous physical activity (MVPA):

“Activity trackers appear to be effective at increasing physical activity in a variety of age groups and clinical and non-clinical populations. The benefit is clinically important and is sustained over time. Based on the studies evaluated, there is sufficient evidence to recommend the use of activity trackers.” (Ferguson et al, Lancet Digital Health 2022)

Activity trackers

• Apple Watch uses activity rings and reminders to nudge users toward daily goals (it has a new AI-powered “Workout Buddy”, too, but that focuses on exercise, rather than on regular daily activities). Data from the Apple Heart and Movement study of 160,000 users confirmed that closing rings and related nudges increased movement and reduced sedentary time. [Note: I am a happy Apple stockholder of long duration]

• Fitbit (now part of Google) has also been shown to produce an increase in steps, MVPA, and a decrease in weight.

• Garmin Connect occupies the same category of mainstream trackers as Apple Watch and Fitbit, and likewise has mounting evidence for increasing activity.

• Oura Ring extends coaching into sleep and recovery. While I was unable to find any research that Oura’s version of tracking and coaching increased physical activity (I’m very open to correction on this) there is no reason to believe that it would not produce similar results to the other devices.

• WHOOP, maker of another wrist-worn device, has published research showing that more frequent WHOOP wear was associated with lower resting heart rate (RHR), higher HRV, longer sleep duration and increased physical activity. 

There are, of course, many other examples in this category: more rings, more watches, more trackers.

Beyond activity trackers

There are exciting developments happening in sleep tracking, and in metabolic guidance, from companies like Eight Sleep and Lumen. I’ll detail these in a future post.

Closing the action gap

In Part 1, I described the interpretation gap — the distance between mountains of raw data and scalable actionable insight. The next challenge is the action gap: turning that insight into sustained activity and habit. As an example, Apple’s data shows that their 'Time to Stand' watch reminders reduce sedentary behavior by 12%.

All the products mentioned above address this action gap through behavioral design rather than medical authority or public health regulation. They deliver timely, specific, and continuous feedback—something healthcare has never been structured to provide.

Timeliness keeps users engaged. Specificity makes action clear. Continuity sustains momentum. Each of these supports the broader cycle that defines digital coaching: measure, interpret, act, and repeat.

What the evidence shows—and what it doesn’t

To summarize, the evidence from these device-software combinations now supports several clear points:

• Behavior change is real. Randomized and observational studies show sustained increases in daily activity and improved sleep when coaching features are active.

• Physiologic benefits follow. Users show lower resting heart rate, higher HRV, and better aerobic capacity—objective markers of improved cardiovascular health.

• Clinical endpoints remain unproven. Whether these preventive gains translate into lower disease incidence or mortality will require longer follow-up.

But the early signal is unmistakable: selfcare can produce the measurable improvements that healthcare has always wanted but struggled to deliver.

The economics of selfcare

For the companies building these tools, prevention isn’t charity or a government function — it’s a business. And a simpler one than healthcare ever allowed.

No billing codes. No licensing fights. No waiting for a CMS rule change that never comes. A $200 device or a $10/month subscription delivers the kind of daily guidance that doctors wish they had time for—and that no insurer will pay them to give.

The market is the part healthcare never served: the well, the worried-well, the “I should probably move more” crowd. Doctors see the sick. Prevention sees everyone else, every day.

This is classic Clayton Christensen new-market disruption. These coaches aren’t taking patients away from doctors—they’re turning people who never had a coach into customers. Most of us will never pay $150 for a trainer to tell us to stand up, once. But we’ll pay $250 for the watch that does it at 2 PM, every single day.

The flywheel is quiet but real: more movement → better data → sharper advice → more movement. Apple closes ten billion activity rings a day. Fitbit counts a hundred trillion steps a year. No health system can match that cadence without spending itself into the ground.

Prevention, after decades of slogans, finally has a business model that works.

Healthcare’s death of a thousand cuts

Above, I said that these coaches aren’t “directly taking customers from healthcare incumbents” — and that’s true. But indirectly, they pose a distinct threat to healthcare incumbents in the same way that the GLP-1 drugs like Ozempic do: by making people healthier and reducing their need for care.

Down the line, every extra thousand steps is one less stress test, one fewer statin, one less procedure. A slow, steady shift: healthier people, smaller healthcare bills, and revenue that flows to the device in your pocket instead of the clinic down the street.

Prevention, after decades of talk, finally has both a delivery mechanism and a business model. And every success story outside healthcare makes healthcare itself a little smaller.

The pattern continues

We’ve seen this progression before:

Knowledge moved out of institutions with search.

Measurement moved out with wearables.

Interpretation moved out with AI.

Now behavioral change — the essence of prevention — is moving out through digital coaches.

Each step makes individuals more capable on their own and shifts both activity and revenue away from healthcare and toward Doctor You. And the technologies transforming health aren’t making healthcare better. They’re making people better—so they need healthcare less.

That’s the trajectory of selfcare, and the quiet revolution of digital coaches.

I’ve written before that the story of modern medicine and modern technology is increasingly a story of shrinking healthcare. Each wave of technology moves one more thing that used to require a professional into the consumer domain — or eliminates the need for it altogether:

• Search engines, and now LLMs give patients the ability to self-diagnose, and potentially self-treat, before ever calling a doctor.

• Smartphones and wearables turned into continuous health data monitors, logging everything from movement to language use to falls and gait steadiness.

• Self-Driving cars aren’t labeled as medical devices, but they’re already reducing the need for trauma care.

None of these started within healthcare, but each is in a real sense competing with it. This is disruptive innovation in Clayton Christensen’s original sense — and the outcome is a smaller healthcare system, not a bigger one.

The newest turn in this process: digital health coaches.

The interpretation gap

When the iPhone and early wearables appeared, they unleashed an ocean of personal data: steps, sleep, heart rate, oxygen levels, even EKGs. Suddenly, ordinary people were collecting data that previously required a visit to a clinic or lab.

But data isn't understanding. Most people couldn't interpret what their graphs meant or what to do about them. And while healthcare providers have the expertise to make sense of that data, they have no framework or tools to bring it into their systems — and definitely no time to review it during a 10-minute office visit.

So for years, we've had this strange paradox: more health data than ever before, flowing continuously from millions of devices, with almost no one equipped to turn it into action. The data existed. The expertise existed. But they couldn't connect.

That gap — between continuous measurement and continuous guidance — is what AI is now closing.

From graphs to guidance

AI has this potential because at a fundamental level, the function of AI is to automate expertise.

On the health side, that expertise might be the reasoning that lets your internist recognize pneumonia from a stethoscope exam, or the experience that helps a diabetes educator adjust insulin doses, or the pattern recognition that guides a trainer in adjusting a workout plan.

Generative AI — the newest and most visible form — extends this by communicating that expertise in plain language, personalized to you, and in whatever language you speak. It can translate professional judgment into accessible advice, continuously and at scale.

From expertise to daily coaching

It’s easy to imagine “digital coaches” only as products for athletes and early adopters (and I’ll talk about common consumer coaches in a future post). But that’s no longer true. Increasingly, these systems are performing core healthcare tasks: monitoring vital signs, interpreting results, adjusting treatment, and recommending next steps — functions that were once the exclusive domain of doctors and nurses.

A few examples:

• Dexcom, best known for its continuous glucose monitors, has recently launched a generative AI platform that builds personalized guidance on top of real-time glucose data. Combined with partners like January AI, these systems analyze thousands of data points each day to prompt insulin users on when to eat, dose, or exercise — automating a large part of diabetes management that once required professional oversight.

• Teladoc Health, including the formerly-independent Livongo platform for diabetes management, connects glucose meters with AI-driven analysis and messaging. It reviews each reading in real time, recommends behavioral or medication adjustments, and alerts clinicians only when values become dangerous. The daily interpretation and medication titration that once required endocrinology visits now happens automatically, at home, for hundreds of thousands of patients.

• Omada Health, which filed for an IPO in 2025, extends that model to hypertension, diabetes prevention, and weight management. Using connected blood-pressure cuffs, scales, and continuous engagement through its app, Omada’s system flags concerning trends and adapts guidance between visits. What used to be a doctor’s or nurse’s job — monitoring, advising, adjusting — has been absorbed by software and remote coaches.

Each of these systems moves core medical interpretation — reading results, identifying patterns, advising action — out of the clinic and into the consumer domain. They aren’t about optimizing athletic performance; they’re about replacing routine clinical decision-making with continuous, automated guidance.

Disruption that shrinks healthcare

Christensen described how disruption begins with "non-consumers" — people not currently served by incumbents — and then advances until it competes head-on.

Digital health coaches started with people we don't usually think of as underserved: people between doctor visits, managing conditions on their own, or simply trying to make sense of the data their devices produce.

Those people may see a doctor once a year, or even three or four times a year — but does that really constitute adequate "service," particularly for those with chronic disease? Another way to think of it: if we could have a physician visiting each of these people every day at home, monitoring trends and adjusting guidance in real time, would outcomes improve?

Well, we're about to find out, because these technologies now make that possible — without the physician.

Digital health coaches aren't improving access to doctors. They're replacing the interpretive and advisory work doctors do. They're moving clinical judgment — the daily monitoring, pattern recognition, and decision-making that used to require years of training — into software that runs continuously, at scale, at near-zero marginal cost.

For healthcare organizations, the strategic question isn't whether to adopt these tools. It's whether the work your clinicians do today will still require a clinician tomorrow — and if not, what you'll do instead.

The most transformative technologies don't make healthcare better.

They make us need it less.

That's not a threat to resist. It's a reality to navigate. The organizations that thrive will be those that recognize which clinical tasks are moving out of their walls, partner with or build the platforms that automate them, and refocus their talent on the complex, high-stakes work that still requires human judgment.

The question for every healthcare executive reading this: what is your organization doing today that software will do tomorrow — and what will your organization do tomorrow that software can't?

In a future post, I’ll explore how consumer-focused digital coaches are reshaping wellness for everyone—not just patients.

After the U.S. Senate removed a proposed national moratorium that would have blocked states from regulating AI, 29 legislatures introduced 130+ health-AI bills, many targeting claim denials and prior authorization.

More specifically, legislators are trying to prevent insurers from letting AI deny claims without the review of a physician. A good example is Arizona's H2175, which states that AI “may not be used to deny a claim or a prior authorization” and that “a health care provider shall individually review each claim” if medical judgment is involved.

This process has been remarkably bipartisan: Arizona's bill passed unanimously in both houses, with roughly equal numbers of Democrats and Republicans supporting.

Press coverage, too, has been consistently supportive. As an example, a recent NBC News article on Arizona's bill provides supportive statements from the Arizona Medical Association, the sponsoring lawmaker, the American Medical Association, and a Harvard Law School professor. From the Arizona Medical Association:

Across NBC, PBS, Fox, and local outlets, coverage was uniformly supportive — with not a single dissenting view. The unanimity is striking — perhaps the first time PBS and Fox News have found themselves on the same side of a health policy debate.

But I get this overwhelming support, I do. On the surface a universal human sign-off seems compassionate and smart. But the effort is misguided, and risks prolonging the same problems that make the existing, non-AI systems a real problem for patients, doctors, and insurers.

I understand why these laws exist

Before I argue against these bills, let me be clear: the concerns driving them are legitimate. Several current lawsuits suggest that insurers have deployed AI systems badly, and that patients have been harmed.

The most notorious example came to light in 2023, when a lawsuit revealed that UnitedHealthcare was using an AI algorithm called nH Predict to determine how long elderly patients should remain in post-acute care facilities. According to the complaint, the system was known to have a staggering 90% error rate, yet the insurer continued denying claims based on its output, knowing that only a tiny fraction of patients would contest the denial. Patients were allegedly forced to leave rehabilitation facilities before they were ready, or their families faced crushing out-of-pocket costs to continue care. And in February of this year, a federal judge allowed the lawsuit to proceed.

Similarly, Cigna is currently being sued for using an automated system called PXDX that allowed doctors to deny claims in bulk---reportedly reviewing cases at a rate of one every two seconds. While Cigna claimed doctors were making the final decisions, the speed pretty convincingly suggests the "review" was perfunctory at best.

These aren't hypothetical risks. If supported by the courts, they represent real failures that may have caused real harm. And they explain why lawmakers---and the public---are skeptical of letting insurers use AI without mandatory human oversight.

But prohibition just brings us back to square one

Here's where I part ways with the legislative response: these alleged failures appear to have been the result of bad faith, poor implementation, and inadequate oversight — not proof that AI is inherently unfit for coverage decisions. In both cases above, the problem wasn't that AI made the determination---it was that faulty systems were deployed without transparency or meaningful accountability. 

Banning AI-only denials doesn't fix any of that. It just forces insurers to add a human signature to the process — a signature that, as the Cigna example shows, can be just as perfunctory and meaningless as an automated one. And we'll still have the same lack of transparency about how decisions are made, the same difficulty appealing denials, the same abysmal speeds, and the same high cost.

In other words, banning AI-only denials condemns us to the existing system that everyone already hates.

If we truly want to improve the system, the question isn’t whether to use AI, but how

Frustrations with the existing human-mediated claims systems come down to

1. Speed — The universal frustration. Claims decisions take too long, slowing down care, disrupting doctors’ workflow, and delaying payment.

2. Transparency — Denial letters are vague, with unclear reasons for denial, making it impossible to know if the decision is medically sound.

3. Cost — The insurer’s primary concern. Manual processes and human review are expensive, and those costs ripple through the system to premiums and taxpayers.

What the legislators (and the media) are missing is that we can use well-designed AI to address all three issues.

Reviewing medical necessity determinations is a narrow, highly-structured task: reading documentation, checking it against coverage rules and practice standards. Modern LLMs already handle similar structured tasks in finance and logistics — and can do the same for claims.

An AI system can also promote transparency. Automated systems are better than people at producing plain-language, structured, reproducible rationales, and they can do it 24/7/365. They can log which policy clause applied, what evidence was missing, and how the decision could be overturned. That's far more detail than a cursory physician denial letter. The solution isn't to remove AI; it's to specify the outputs we want  and require AI systems to provide them.

Further, we can engineer quality into these systems much more easily and cheaply than with people. We can build-in audits, sampling, and benchmarks, that ensure verifiable fairness and reproducibility. We can also create HIPAA-compliant ways to allow third-party audits including human audit, but probably AI reviewing AI.

And this last point isn't theoretical: existing tools like Claimable, Fight Health Insurance, and the completely free Counterforce Health already exist to help patients generate appeal letters and contest denials in minutes. The Guardian has described this as an emerging "AI vs AI" arms race, where patients' tools counter insurers' tools. The key is to remove barriers to using these oversight AIs by standardizing access to decision packets and rationale and allowing these developers to battle it out to improve their products.

A better path: automation with the right guardrails 

The future of our slow, opaque, and costly system shouldn’t be to burden doctors with more administrative tasks that pull them from patient care. Modern LLMs can handle the narrow, bureaucratic task of reviewing claims against coverage rules. By requiring AI to provide plain-language denial reasons, instantly accessible to doctors and patients, and enforcing continuous audits by regulators, we can ensure fairness and transparency. This lets doctors focus on care, patients get clear answers, and insurers cut costs — a win for all.

This is no different from how we’ve approached autonomous vehicles. Few argue we should ban self-driving cars outright — because that would eliminate the chance to improve a technology with huge potential to save lives. Instead, we regulate and monitor them, creating safety benchmarks and oversight while allowing the technology to evolve. Insurance coverage decisions deserve the same approach: guardrails, not prohibition.

Who will be first with instant, contestable, transparent denials?

Banning AI-only denials feels humane, but it entrenches the worst features of the old system: delay, opacity, and cost. The real solution is not to put more doctors into the administrative loop, but to design AI systems that are fast, transparent, and auditable  and to guarantee that patients and doctors have the documentation and tools needed to appeal effectively, including with AI.

Patients don't need slower denials with a human rubber stamp. They need faster, clearer, and more contestable decisions — and for the first time, AI makes that achievable. The first insurer to deliver instant, contestable, transparent denials will win trust and market advantage — and regulators should be enabling that, not blocking it.

When the naturalist John Muir penned that quote, he was writing about the natural world after working as a shepherd in California’s Sierra Nevada, immersed in the mutual dependence of mountains, forests, rivers, and animals.

More than a century later, his insight is proving just as true for the wilderness inside of us. We’re discovering that nothing in our physiology exists in isolation. The eye, the heart, the skin, the voice — each carries hidden connections to other systems. And increasingly, it is AI that is uncovering those links.

An Early Example: Apple Heart Study

In 2017, Apple and Stanford launched the Apple Heart Study, enrolling more than 400,000 participants. Researchers asked a simple question: could the Apple Watch’s pulse sensor — a photoplethysmograph (PPG) LED originally designed just to measure heart rate — be used to detect something far more serious?

The answer was yes. By applying machine learning, Apple showed that irregular pulse patterns captured by PPG could flag atrial fibrillation (AFib), a condition often silent until it causes a stroke. The connection between a consumer watch’s pulse sensor and a dangerous cardiac arrhythmia was a hidden connection — one not previously recognized or used at scale. And gaining the valuable additional information didn’t involve any new hardware at all.

More Hidden Connections Revealed by AI

Since the Apple study, machine learning has surfaced hidden links across medicine, and not just with watches and phones and other mobile devices:

• Retinal images — long used to check eye health — can reveal sex, race, smoking status, blood pressure, and even cardiovascular risk.

• Gait and voice patterns — captured casually by smartphones or microphones — can detect the earliest signs of neurological disease without a visit to a neurologist

• Electrocardiograms (ECGs) — the same tracings used for decades to diagnose heart rhythm — now allow AI to screen non-invasively for anemia, electrolyte imbalance, and left ventricular dysfunction

• CT scans — often ordered for one purpose — can be re-analyzed at minimal cost to estimate bone density or coronary calcium, each tied to long-term risk — without needing additional scans (or radiation)

The information was always there. What’s changed is that AI gives us the ability to see it.

The Apple Watch and Hypertension

This is particularly true for Apple’s new hypertension notifications, because they don’t use any kind of new sensor in the Apple Watch. The same green LEDs still measure blood oxygen levels. The electrodes still capture heart rhythm. The accelerometer still tracks movement. What changed was Apple’s recognition that these existing signals already carried the imprint of something new: hypertension.

For years, millions of Apple Watch users have been streaming this data. But it’s only now — after training AI on the right relationships — that the watch can flag blood pressure risk with confidence.

Another hidden connection uncovered.

Beyond the Watch

We can expect that this pattern will repeat itself everywhere: applying AI to old data streams to produce new insights. Combined, of course, with new data streams — very likely including non-invasive blood glucose monitoring in the near future.

• Credit card purchases — will allow us to understand not just spending, but nutrition and mental health

• Browsing histories — can understand cognition and promote education

• Medical images — will be repurposed by AI to surface cardiovascular, skeletal, or metabolic risk factors unrelated to the original reason for the scan

Muir’s observation was ecological, but it applies just as well here. Health is not a set of isolated numbers. It is a network of hidden links — and we are only beginning to trace them.

Closing Thought

Apple’s hypertension notifications are not the end of something — they are the beginning. They can help us imagine how every stream of data we collect may reveal more hidden connections between body systems than we ever imagined. And the future of health will be built on our ability to recognize those links connecting what we already measure to what we most need to know.

What I hadn’t considered was sustainability, the environmental cost buried in our healthcare system, and how that is affected by consumerization. But recently I was prepping for a conversation with someone working on sustainability at a pharma firm, and it opened my eyes.

Houston, We’ve Got A (Environmental) Problem

Healthcare is a bigger polluter than most people realize, and — surprise! — emits more carbon globally than the shipping industry (Think Global Health).

• In the U.S., it’s responsible for roughly 8.5% of the greenhouse gas emissions (Commonwealth Fund).

• More than 70% of that footprint comes from indirect sources—things like the supply chain for drugs and devices, and patient and staff travel (AAMC).

• Medical waste is staggering: the U.S. health system generates about 6,000 tons daily, much of it single-use plastics.

• Inhalers alone are a big contributor. In 2022, inhalers used by Medicare Part D and Medicaid beneficiaries generated emissions equal to the annual electricity use in Riverside, California or Buffalo, NY (PMC).

So despite its many benefits, the current system of healthcare also has a gigantic negative environmental impact. But consumerization works against that — even just by reducing the number of times people need to drive to the doctor’s office.

Think about it: healthcare’s carbon footprint in half would save more than:

• turning Florida carbon neutral

• grounding every US airplane

• making every American home zero-carbon

🤯

Fewer Visits, Less Pollution

Think about what happens when telemedicine — enabled by consumerization-provided high-speed internet and ubiquitous computing devices — replaces clinic visits. A 10-minute video call instead of a 2-hour round trip journey isn’t just more convenient; it cuts emissions by orders of magnitude (even accounting for the impact of electricity generation and transmission, and device manufacture).

As an example, at Stanford Health Care, adding telemedicine cut emissions by 36%, saving 17,000 metric tons of CO₂ over two years (PMC).

It’s the same for the increasing use of online tools and apps, combined with the growing range of over-the-counter medications: fewer visits (for more on this, see In the Future, You’ll Need Your Doctor Less).

We’re also increasingly finding the utility of AI-enabled tools for prevention and early detection (think Apple Watch for atrial fibrillation or deterioration in walking steadiness). And this is just getting started, with enormous potential to reduce hospitalizations—the single largest environmental sink in healthcare.

But It’s Not Just Visits

Of course, reducing clinic visits isn’t the only path to environmental benefit, and pharma companies are working on new approaches that may entail the same number of visits but still way fewer emissions.

One example is GSK’s new albuterol inhaler. Traditional metered-dose inhalers use hydrofluoroalkane propellants, which are powerful greenhouse gases. In fact, inhalers account for about 3% of the entire NHS carbon footprint in the U.K. (Nature). GSK is developing a low-carbon version of their Ventolin (albuterol) inhaler that uses a new propellant with a footprint up to 90% lower than current inhalers. If approved, it could cut GSK’s overall carbon emissions by more than 4 million tons annually (GSK, FT). And with AstraZeneca pursuing a similar product, low-carbon inhalers are likely to become the new standard.

Another example is long-acting HIV injectables. Right now 30 million people worldwide take daily HIV pills. Switching to long-acting injectables means less raw material and less energy to manufacture and transport medications. That’s a huge supply-chain environmental win.

To be fair, neither of these examples is really a consumerization story, at least not yet: patients already use inhalers and pills at home, and injectables often still require clinic visits. But both are moving in that direction. Inhalers, once reformulated, are obvious candidates for eventual OTC status. And HIV therapies are now being tested for safe home self-administration (EATG).

Once home self-administration becomes routine and evidence accumulates on safety and adherence, it’s not hard to imagine both therapies moving over-the-counter. That’s the point where sustainability and consumerization go hand in hand: fewer visits, lower emissions, and more control in the hands of patients.

Healthier and Greener

This all began with my own coming to see the environmental angle of consumerization — though only after someone from pharma raised it. And now I can’t un-see it: healthcare has an environmental problem. Happily, consumerization is already pushing more care away from clinics and into the hands of individuals. Telemedicine, AI-powered wearables, and at-home monitoring show how fewer visits mean fewer emissions—and they’re proof that consumerization and sustainability already go hand in hand.

What’s changing now is the scope of that connection. Product innovations like low-carbon inhalers and long-acting HIV therapies start as sustainability wins, but they only become true consumerization stories once patients can use them independently — ideally through safe home self-administration or eventual OTC access. Taken together, these trends point to a future where consumer-driven health doesn’t just make care more convenient and more effective, but also measurably greener.

“What does it mean for the future of health when Apple makes more money from AirPods than the top ten EHR vendors combined?”

At the time, it might have seemed like an unfair comparison—enterprise software versus global consumer electronics. But five years later, it’s clear that this contrast was not only valid, but also predictive of a broader shift: health data—and thus health innovation—has moved out of the clinic and into the consumer’s pocket.

The Real Center of Health Data Is Not the EHR

The Electronic Health Record (EHR) remains the data centerpiece of traditional medicine. But in terms of data volume, it’s no longer the main source of information about our health. And it’s not even close.

U.S. EHR Data Volume

Each U.S. patient generates an estimated 80 megabytes of EHR data annually, including clinical notes, lab results, and imaging. Multiply that by roughly 250 million Americans interacting with the healthcare system annually, and you get:

→ 20 petabytes per year of U.S. EHR data

U.S. Consumer Health Data Volume - Calculated by Total World Data

There is no question that the real volume, and the real growth, in health-related data is happening outside the exam room, and outside the EHR.

Every day, we generate health-related data through wearables, smartphones, browsers, keyboards, credit cards, sleep monitors, voice assistants, and fitness apps. Not to mention smartwatches, rings, and earbuds.

One way to estimate this is by looking at the total data produced worldwide:

Global data generation reached approximately 147 zettabytes in 2023. If we very conservatively estimate that only 5% of this is health-related consumer data (i.e. somehow related to, or useful for, our health), we get approximately 7.4 ZB/year globally.

Adjusting for the U.S. — which has about 4-5% of world population but likely 10-15% share of digital activity — we can conservatively estimate that the U.S. generates roughly 1 zettabyte (1,000,000 petabytes) of consumer health data annually..

→ That’s 50,000 times the volume of U.S. EHR data

Another way is to look at individual data generated.

For example, everyone with a smartphone — which Pew estimates to be 91% of the roughly 230 million US adults, or 209,000,000 people. Even if we just look at the accelerometer data for those people, leaving out all the other health-related information on their phones (like speech patterns, word choices, grocery bills, etc) continuous low-frequency sampling (1–2 Hz) on an iPhone can generate 1 GB of raw movement data annually. Which is 209 pettabytes for the adult population.

→ That’s still ten times the volume of U.S. EHR data

So even our low-end estimate shows that we’ve got 90%+ of our health data sitting on our phones, not in the EHR. And it’s not hard to believe that if we wanted to predict, for example, cardiovascular risk, using either the average person’s EHR data (which is likely minimal) versus that person’s continuous data on how fast, how far, and how often they move . . . it’s the phone data that will produce a better prediction.

Although as far as I know, no one is yet doing that. Sigh.

Innovation Follows Data

If the EHR holds such a tiny part of U.S. health-relevant data, where should we expect most of the next breakthroughs in diagnostics, prediction, and prevention to occur?

Not in the hospital. Not in the clinic. Not in the EHR.

Innovation is flowing to where the data is—in consumer-facing platforms and tools. We’re already seeing this play out in:

• AI Symptom Checkers: Platforms like K Health and ChatGPT-based tools provide users with AI-driven assessments of their symptoms, offering preliminary insights before consulting healthcare professionals.

• Wearable-Based Sleep Optimization and Stress Tracking: Devices such as the Oura Ring and WHOOP monitor sleep patterns and stress levels, providing users with actionable insights to improve rest and manage stress effectively.

• Early Illness Detection: Wearables like the Oura Ring have introduced features such as Symptom Radar, which can detect early signs of respiratory illnesses by analyzing physiological data, enabling users to take precautionary measures promptly.

• Fitness and Glucose Feedback Loops: Dexcom G7 offers continuous glucose monitoring that now connects directly to Apple Watch, allowing users to track their glucose levels in real-time. Similarly, Levels and Ultrahuman provide platforms that integrate glucose data with lifestyle factors to optimize metabolic health.

• Personalized Coaching and Risk Scoring: Apps like Lark Health utilize AI to offer personalized health coaching, helping users manage conditions such as diabetes and hypertension through tailored advice and monitoring.

• Mood Inference and Mental Health Support: Ellipsis Health has developed voice-based technology that analyzes speech patterns to assess mental health, providing insights into conditions like depression and anxiety.

These advances are possible not because of structured ICD codes—but because of rich, continuous consumer data at massive scale.

Conclusion: The Real Medical Record Is Distributed

In 2025, the EHR remains essential—but it’s nothing like central. The richest, most frequent, most predictive health data lives outside the clinic.

And that means the future of health won’t be driven by Epic, or Cerner, or Allscripts. It won’t even be driven by United Healthcare, or Kaiser, or Mass General.

It will be driven by whoever best understands—and leverages—the flood of health-relevant data being generated by consumers, not patients. And that is likely to be the consumer tech companies, both giants like Google and tiny startups or that haven’t even left the garage yet.

Sources

• Forbes – “The Skyrocketing Volume of Healthcare Data”

• Google Blog – “Pixel’s AI-Powered Cough Detection”

• Exploding Topics – “How Much Data Is Generated Per Day”

• Wikipedia – “AirPods Sales Volume”

• Journal Article - “Smartphone-Based Lifelogging: An Investigation of Data Volume Generation Strength of Smartphone Sensors”

Let me correct that with this post.

In 2024, U.S. traffic deaths dropped nearly 4%, the sharpest single-year decline in half a decade. The National Highway Traffic Safety Administration (NHTSA) credits this trend in part to technologies like automatic emergency braking (AEB), which will be mandatory on all new light vehicles by 2029. NHTSA estimates that AEB could prevent 360 deaths and 24,000 injuries annually once fully deployed.

That’s before you even factor in what’s coming next: widespread autonomous driving. Tesla’s Autopilot — definitely not yet a fully autonomous driving system — already reports one crash for every 7.4 million miles: seven times safer than the U.S. fleet average. Waymo’s driverless fleet in Phoenix has now driven tens of millions of miles with zero at-fault fatalities. These systems are getting better, quickly. They don’t speed. They don’t text. And they don’t fall asleep at the wheel.

And even now, before full autonomy is widespread, the downstream effects of safer driving technologies are already being felt—just not yet on hospital dashboards.

For 2019–2020, an average of 3.8 million emergency department (ED) visits for motor vehicle crash injuries occurred annually — about 3% of ED visits. And car crashes are not low-cost visits. They trigger trauma activations, CT scans, orthopedic surgeries, ICU stays, and weeks or months of rehabilitation.

In 2022, motor vehicle accidents in the US resulted in over $470 billion in total costs, including medical costs and the cost of lost lives, according to the CDC.

Which is exactly why a slow decline in motor vehicle trauma should be setting off alarms in every hospital boardroom.

Trauma, orthopedics, imaging, rehab—many of healthcare’s most dependable, well-reimbursed services depend not just on illness, but on accidents. And no single type of accident has been more reliable, more predictable, and more lucrative than the car crash.

Autonomy changes that.

Not overnight. But incrementally, and often invisibly. There’s no press release saying “trauma cases will drop 15% this year.” There’s just a software update. And then another. And another. Every version pushes trauma volumes a little lower.

If you run a trauma center, your real competitor might not be the hospital across town—it might be the software quietly reducing your patient volume from the outside.

And it is a major threat. Trauma centers are high fixed-cost operations. They can’t scale down neatly when volume dips. You still need the surgeon on call, the CT scanner, the blood bank, the full staff. When volume becomes unpredictable, the economics start to break.

Healthcare tends to expect disruption to look like a healthcare competitor—CVS Health, Walmart, Amazon Clinic. But true disruption rarely comes from within the market. Blackberry didn’t fall to another mobile phone company, it fell to Apple — a company that had never made a phone. Taxi companies didn’t fall to another transport company. They fell to software layered onto private cars.

Healthcare’s trauma volumes may fall to something even simpler: safer roads. And any strategic plan built around growing orthopedics, expanding trauma capacity, or maximizing downstream imaging needs to be re-examined through the lens of this shift.

What if car crashes drop 20% in five years — from software updates?

If you’re a healthcare executive, or board member, you might not need need to change your mission. But you do need to change your assumptions. You can’t control the crash rate — but you can control how prepared your system is for the rapidly-approaching world in which crashes are rare.

References:

• NHTSA Final Rule on Automatic Emergency Braking (AEB)

• Tesla Autopilot Crash Data Q1 2024

• Waymo Safety Impact Report

• Overview of Emergency Department Visits Related to Injuries, by Cause of Injury, 2017

• Emergency Department Visit Rates for Motor Vehicle Crashes by Selected Characteristics: United States, 2019–2020

And this didn’t begin with ChatGPT. For decades, consumers have used online search engines, wearables, and over-the-counter medications to bypass the traditional gatekeepers of care. AI simply supercharges a longstanding trend: the democratization of health knowledge and decision-making.

What happens when free, always-on tools help patients adjust their insulin, distinguish between a cold and pneumonia, or monitor key metrics continuously rather than once a year? The answer isn’t a dystopian future without doctors—but it is a future where clinicians must reorient their value toward what cannot be automated.

In the piece, I encourage physicians to embrace AI, not fear it—to find new roles, contribute to the development of patient-centered tools, and show the unique value only human clinicians can provide.

If you’re a health leader, technologist, or clinician, I’d love to hear how you’re navigating this shift. The question is no longer whether change is coming—but whether doctors play the role of leader or observer.

🔗 Read the full article in American Family Physician:

https://www.aafp.org/pubs/afp/issues/2025/0300/editorial-artificial-intelligence-consumer-empowerment.html

Most discussions of disruption in healthcare assume the healthcare system remains essential — just reshaped. Virtual visits replace office visits. AI handles documentation. Automation reduces friction. But the assumption is always that the care is still needed.

What happens when it isn’t?

Some of the most significant disruptions now emerging aren’t aimed at healthcare, but may make large parts of it unnecessary.

Prevention That Sidesteps the System

For example:

• GLP-1 drugs may reduce obesity, diabetes, and hypertension across entire populations. In fact, in 2023 the US population obesity rate declined for the first time in a decade — and one leading theory attributes it to the effect of GLP drugs.

• Self-driving cars could cut motor vehicle trauma by 30% or more. RAND estimated back in 2017 that even a 10% decrease in accident rate could save “hundreds of thousands of lives”. Meanwhile, even current technology promises a much greater reduction, with Waymo reporting a 57% decrease in police-reported accidents compared to human drivers.

• AI therapy chatbots are already showing promise for mental health treatment, and could dramatically lower the cost of therapy — and might prevent mental health crises before they begin. And they may just be the thin edge of the wedge, presaging the automation of all medical conversations where only information is exchanged.

Endocrine Disruptors

Each of these innovations reduces the need for healthcare, rather than streamlining its delivery or improving its process. And yet few health systems — actually, none that I am aware of — are modeling what happens when demand drops dramatically, not for elective procedures but for entire categories of care.

Food companies have run scenarios to prepare for GLP-1-driven shifts in consumption, and are already rolling out products tailored to GLP users. Where are the equivalent conversations inside healthcare?

If the number of diabetic patients drops by 20% (a conservative estimate, given the effectiveness of GLP-1 agonists like Ozempic), what happens to the business model of the average hospital? Will we need fewer endocrinologists — or none at all? What happens to primary care visits (already drastically declining) if hypertension declines by a similar percentage? What happens to orthopedics and emergency medicine if autonomous vehicles mean car crashes plummet by 50? Who owns prevention — and what happens to the institutions built on managing failure to prevent?

Failing to Plan is…

Healthcare leaders need to consider, in a world where health problems are increasingly prevented, not treated — what’s left for the system to do? I’ve spoken with hundreds of healthcare leaders in the past five years — including dozens of CEOs — and not one has mentioned the kind of “best case / worst case” modeling now common in the snack food industry.

If healthcare is going to meet this moment, it needs to do better.

And yes, there’s something ironic about hospitals taking planning cues from potato chip companies. But maybe that’s exactly what needs to happen.

Much of this post has focused on technologies that reduce the need for healthcare by preventing illness or injury. But the most widely discussed disruption today — generative AI — may reduce the need by eroding healthcare’s monopoly on information and expertise.

From diagnostic chatbots to at-home guidance systems, AI is already starting to displace parts of the care experience.

I’ve written about that in previous posts — and I’ll return to it soon.

• Part I, introduced the idea of “disruption” as first defined by the late Clayton Christensen in his 1997 best-seller The Innovator’s Dilemma: the process whereby a new technology creates a new business model, allowing new companies to successfully overturn incumbents.

• Part II provided specific examples of disruption within the healthcare system: from Walmart Health to IDx-DR, the first AI (artificial intelligence) system approved by the FDA to make a medical diagnosis (of diabetic retinopathy) without a doctor’s input, to Chestlink , an AI system designed to read chest x-rays.

This post focuses on even bigger disruptions to the whole healthcare system, on innovations that move things not just from a specialist to a generalist (like IDx-DR), or from a hospital clinic to a Walmart clinic, but away from providers and clinics entirely.

I refer to this as “selfcare”: the system whereby untrained consumers can perform former healthcare tasks without interacting with a “provider” like a doctor or nurse.

I’ll start with over-the-counter (OTC) medicines, a great example of disruptive selfcare that is completely under the radar of most in healthcare. Then I’ll move on to more directly IT-intensive disruptors like AI.

The IT revolution goes consumer

Computing has gone through three major stages over the last 100+ years, when viewed from the eye of the consumer. In the first, which reached all the way up into the 1970s, computers were so expensive and difficult to purchase and maintain that only large institutions (government, industry, universities) could be customers.

The second stage began in the 1970s with the early “personal” computers, and ran all the way through the beginning of the 21st century. In this stage, the cost (and size) of components dropped sufficiently that it was possible to begin selling computers (i.e. PCs) to much smaller institutions, including small businesses or even individuals.

The third stage, which has characterized computing in the 21st century, is the consumer stage: the biggest chunk of the worldwide market for information technology is for consumer technology, with smartphones taking the biggest piece of that chunk.

Compared to the roughly $1.5 trillion projected for consumer IT spending, the portion expected for healthcare, roughly $200 billion, is small potatoes. And these numbers reflect a substantial decrease in consumer spending projected for 2022 after the huge spike in lockdown-related sales of laptops and other computers to consumers during the pandemic.

The augmented consumer

These IT spending numbers matter, because in our lifetimes the most important way that we all gain new capabilities is via IT (especially smartphones). And we’ve all gained so many new capabilities via smartphones and apps that it’s hard to keep track.

In healthcare, as reflected in its laggardly IT spending? Not so much.

The biggest IT spending in healthcare in the last ten years has been to purchase “EMRs” (electronic medical record systems), partly with $30+ billion in federal subsidies. Many hospitals systems have spent upwards of $100 million on their EMR, and the big EMR companies (like Epic and Cerner) are tremendously influential in the healthcare system.

But guess what: consumers spend more money each year on AirPods than healthcare spends each year on EMR! And because of IT, consumers have been gaining new capabilities much more rapidly than the healthcare system.

This wouldn’t matter for healthcare if consumer capabilities were unrelated to healthcare capabilities. But they are very related: many of the things that IT now lets consumers are do for themselves are things that used to be “healthcare.”

You don’t need AI to replace doctors

Over the last few decades, hundreds of medicines have been moved from prescription to OTC. The FDA said in 2016 that there were more than 700 drugs that “would have required a prescription only 20 years ago” and this number has presumably grown since then. Those would include commonly used drugs like Claritin (loratadine) or Advil (ibuprofen) — hard to remember now that you used to have to visit a doctor to take Advil!

Part of the reason why FDA can move drugs to OTC, of course, is that consumers today have many more tools to find out medical information than thirty years ago: computers, the internet, Google, drugs.com, Medscape.com, smartphones, to name just a few. Consumers equipped with Google are much better able to educate themselves about their health.

I’ve been asked many times whether I think that AI will replace doctors. Never once have I been asked if I thought that OTC drugs could replace doctors. But that’s exactly what they have done: every time a drug switches from prescription to OTC, the total number of doctor visits drops.

In fact, a 2012 study by Booz concluded that

“if OTC medicines did not exist, an additional 56,000 medical practitioners would need to work full-time to accommodate the increase in office visits by consumers seeking prescriptions for self-treatable conditions.”

Fifty-six thousand doctors that we don’t need because of OTC drugs; that’s almost 6% of the practicing doctors in America. Think of the effect on healthcare costs.

But AI is an accelerant

If we thought that Google-equipped consumers were better able to manage their health, AI is going to blow that away. A product called ResAppHealth is a powerful example.

We’ve probably all heard something about the uproar in radiology concerning AI. Most of this was sparked by a 2016 comment by eminent AI researcher Geoffrey Hinton saying that new types of AI were going to be better able to read images than expert human radiologists.

So Hinton was really describing the possible disruption within healthcare of substituting an algorithm for a radiologists but leaving the rest of healthcare intact, as in the image below:

ResAppHealth (now a subsidiary of Pfizer) gives us an idea of what Hinton missed, and how AI is more likely to disrupt the whole of healthcare, rather than just a piece here and there.

ResApp was created by Australian research physicians who thought “when I want to know if a patient has lung problems, I listen to their breathing with a stethoscope to detect any pattern characteristic of pneumonia. I wonder if we could automate that process with AI.”

So . . . now they have a smartphone app that listens to your cough and tells you if you have pneumonia. And they are marketing it to healthcare providers and directly to consumers.

Now, imagine the market for a tool that gives every person the ability to diagnose, or rule out, pneumonia (every parent on earth, for starters). And imagine how that affects healthcare:

Just as with OTC drugs, information technology gives consumers a capability they didn’t have before. And suddenly, a skill requiring 10+ years of training becomes a smartphone app. And that aspect of healthcare — reading x-rays to check for pneumonia — suddenly . . . isn’t healthcare anymore. Costs go down dramatically, more people in need of care are able to get it.

Even the in-healthcare examples I’ve provided in Part II, like IDx-DR (the AI system that diagnoses diabetic retinopathy without an ophthalmologist) are not likely to stay in-healthcare very long. Although for business (and regulatory) reasons that company is now marketing the product to primary care practices, there is no technological reason why that machine couldn’t be sitting right next to the automated blood pressure cuff in the drugstore.

What happens now

As Yogi Berra famously said, “it’s hard to make predictions, especially about the future.” Nonetheless, it’s very, very clear at this point that even without AI, information technology is able to shift the balance between the capabilities of the healthcare system and the capabilities of individual consumers.

Twenty years of over-the-counter drug switching by FDA, at least partly enabled by consumers access to information technology, has already dramatically reduced our reliance on healthcare for a variety of conditions, and reduced the number of clinicians needed by US healthcare by something like 5-10%.

As those same consumers gain access to the automated expertise that is AI, there’s no question that this process will accelerate.

Consumers can sit back and enjoy the process.

Healthcare needs to be thinking of how other organizations in other sectors — like news or travel — survived digital disruption, and pulling lessons out of those examples.

And policymakers need to be revising their predictions of “physician shortages” in light of the last thirty years of selfcare, and the promise of the next thirty.

This post draws on themes from my keynote talks.

A few weeks ago, FDA provided new data:

October 19, 2023 update: 171 Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices were added to the list below. Of those newly added to the list, 155 are devices with final decision dates between August 1, 2022, and July 30, 2023, and 16 are devices from prior periods identified through a refinement of methods used to generate this list. 

With this release, I wanted to see if my earlier findings still held, how fast innovation was accelerating, how quickly were more companies getting involved — and what other surprises lay within the data.

Anti-exponential: medical AI innovation is not accelerating

Scoring points: approvals by year

Probably most interest of all findings is that AI innovation in medicine — at least by the measure of AI device approvals — isn’t accelerating at all! The graph below shows good, steady growth in approvals per year (an average 26% increase per year over the last 4 complete years) but not acceleration:

Of course the number of approvals depends on at least two variables: how many applications are submitted, and how many are approved. It’s possible that the rate of invention (i.e. the creation of new algorithms and tools) is accelerating but that the FDA approval process is the rate limiter.

Joining the team: new companies per year

In terms of broadening the group of innovators, the list this year includes 19 entities that previously had not appeared, including stalwarts of diagnostics like Beckman Coulter and Biomerieux:

• AgaMatrix Inc.

• AlgoMedica

• Appian Medical Inc.

• Beckman Coulter, Inc.

• Biomerieux, Inc.

• Bruker Daltonics, Inc.

• ClearView Diagnostics Inc.

• Cydar Ltd.

• Gauss Surgical Inc.

• IRIS Intelligent Retinal Imaging Systems, LLC

• LabStyle Innovations Ltd.

• Matakina Technology Ltd.

• Monarch Medical Technologies

• Pathwork Diagnostics Inc.

• QView Medical, Inc.

• Renalytix AI, Inc.

• Stratoscientific, Inc.

• Tyto Care Ltd.

• Zepmed, LLC

Liked approvals per year, this is an increase but not an acceleration. As shown in this graph:

So neither the number of approvals nor the number of companies winning those approvals is accelerating.

Less lonely at the top

Last year I noticed that approvals were becoming more concentrated — with only 18% of companies on the list gaining more than half of total approvals (this became 20% with revision of the list by the FDA; see notes at bottom). So far in 2023, you’ve got to include 30% of total companies to get to 50% of total approvals — so approvals are going to a broader group. Hard to say from the data whether this is a trend or just variation year to year, and I’ll plan to revisit when next year’s data comes out.

Your grandfather’s medical AI?

Age of the players

I was surprised last year to find that the average age of the top 6 companies in the approvals list (including all years) was 89 years!

The most surprising point hidden in the dataset was that a lot of the action in medical AI isn’t coming from Big Tech companies known for AI like Apple or Amazon or Google/Verily but from much much older Big Med Tech industrial companies, with GE (founded 1892) and Siemens (founded 1847) taking the number 1 and 2 spots, respectively.

That has changed a bit. The top 12 companies for 2022-23 have an average age of 58 years - so we can count that as a major shift towards new players like Annalise-AI (4 years old), Viz.ai (7 years old), Hyperfine (9 years old), and Aidoc (7 years old).

This is also reflected in the continued drop in the median age of companies receiving at least one approval, as shown in the graph below.

The bump in average age for 2023? That’s in part from the addition to the list of Cedars-Sinai Medical Center (founded 1902) and MD Anderson Cancer Center (founded 1941). Will such venerable institutions continue to move onto the list? Given the accumulated medical knowledge there, let’s hope so.

The graph above just looks at the age of companies “in the game,” with at least one approval. But that means that for 2022, for example, it counts GE, with 17 approvals, the same as United Imaging — with just one.

The graph below gives us more insight into that game, showing what percentage of players are “older” (>= 20 years) versus younger.

Points scored by young vs old

It’s also important to know whether older vs younger players are racking up points (i.e. approvals). And here we see that despite the influx of younger companies the older companies are more than holding their own. In fact, as noted last year, the percentage of approvals going to older companies is steady or possibly increasing.

Big tech MIA

Note that although older companies are holding their own, those companies mostly aren’t Big Tech. Since 2017, Apple’s only gotten 2 approvals, Google’s Verily 2, Microsoft 1, and Amazon 0. That’s a total of 5 out of 531 approvals: less than 1%.

What’s it all mean

At least on the medical side, as measured by FDA AI device approvals, AI innovation is chugging along — but not accelerating. This comes as a bit of a useful corrective to the relentless AI hype. It also raises the question as to whether the rate-limiting factor here is the FDA approval process itself (not hard to believe considering the slow pace of drug approvals), or something intrinsic to the market.

More companies are entering the space and getting approvals. Most of these are younger, but the entrance of MD Anderson and Cedars-Sinai stands as an example for the well-established hospital systems; an example which I hope will spur more participation from the healthcare establishment not just in adopting AI innovation but in helping to create it.

The percentage of approvals going to the older companies is, if anything, increasing. This is classic sustaining innovation in the medical device space, as companies like GE add AI components to improve their existing product line. It remains to be seen who in the space will start to significantly outpace the innovation at those older companies and seriously disrupt the space.

Finally, despite great stated and demonstrated interest in health, Big Tech is still not a major presence on FDA’s list. Given the health moves made by Apple, Google, Amazon, Microsoft and others, this is a little surprising — especially on the Apple side, given their focus on device innovation. Will this change, whether through in-house innovation or acquisition? Time will tell.

Notes

1. I’m definitely not an FDA expert, just a doctor and tech guy trying to figure out what’s going on in the world. If you’ve got special insight into any of this, or corrections to the above, please let me know.

2. The FDA’s AI list appears to be incomplete, leaving out at least 4 de novo approvals utilizing AI that I was able to find by cross-checking other FDA references: one from 23andMe, one from Apple, one from Renalytix AI, and one from Viz.ai. Those 4 are in FDA’s searchable approvals database, but not in their AI list. I’ve included those four in the analyses above.

3. The FDA’S list includes this note: “Of those newly added to the list, 155 are devices with final decision dates between August 1, 2022, and July 30, 2023, and 16 are devices from prior periods identified through a refinement of methods used to generate this list.”

4. I was unable to find out when these two companies were founded, and so they’ve been left out of age calculations. If you know, please share!

   • Yukun (Beijing) Technology Co., Ltd, filer of item K213986 for CerebralGo Plus

   • AssembleCircle Corp., filer of item K220903 for WebCeph

In my last post, I introduced the concepts of innovation developed by Clayton Christensen of Harvard. In Christensen’s paradigm, “sustaining” innovation enables incumbent organizations to better please their existing customers, while “disruptive” innovation weakens incumbents by creating new markets with cheaper but arguably worse technology.

As an example of “worse” disruptive technology, I talked about Netflix, which originally sent customers DVDs by postal mail. That seemed ridiculous to many who liked being able to just walk into a Blockbuster video store and then walk out with a video in a short time.

And the incumbents at Blockbuster apparently laughed and laughed . . . but Netflix was able to siphon off Blockbuster customers who were more price-sensitive. Then, when Netflix adopted streaming technology, they took all the remaining customers because at that point they were both cheaper AND faster.

But healthcare is not video streaming, so what about some examples of disruption in healthcare?

Non-IT-based healthcare disruption

There has been a lot of disruption within healthcare in the last decades; with some of it mostly related to information technology (IT) and some not. Let’s look first at some of the non-IT stuff.

Some examples:

• doctors getting disrupted by ancillary providers like nurse practitioners, who have many fewer years of expensive training, and are paid less

• primary care clinics and emergency departments getting disrupted by urgent care centers with a limited menu of services, and without the legacy costs of giant hospital-centric systems

• specialists getting disrupted by other specialists (see next section, below)

Note that, as with our Netflix example, the disrupters are arguably worse in certain ways than the doctors they are trying to replace (sorry, “augment” 🤣). It’s hard to argue, for example, that nurse practitioners are better than primary care doctors. The important thing, however, is that they are good enough for some section of the patient population — and that they are cheaper.

Also, although these disruptive innovations aren’t mostly related to IT, they are dependent on the IT innovations of the last 40 years: it wouldn’t be so easy to run an efficient urgent care center without inexpensive personal computers, etc.

Example of non-IT healthcare disruption: heart disease treatment

Back in 2000, in an article entitled “Will Disruptive Innovations Cure Health Care?” Christensen himself provided an example of healthcare disruption. He and his coauthors talk about coronary angioplasty (or PCI, for “percutaneous intervention”), which enabled cardiologists to treat heart disease patients by inserting tiny tubes called stents into blood vessels — instead of requiring cardiac surgeons to perform cardiac bypass grafts (CABG). This is disruptive technology, for sure, and it’s a good illustration that disruptive technology doesn’t have to be information technology.

PCI didn’t eliminate heart surgeons, of course, but it did drastically reduce their activities: between 2001 and 2008, the annual rate of CABG in the US went down by 30%! So the incumbents (cardiac surgeons) lost a lot of revenue to the disruptive practitioners (interventional cardiologists) doing the PCI.

Keep in mind that CABG operations were, and are, better in certain circumstances but the disruptive PCI was generally cheaper because cardiologists are cheaper than cardiac surgeons, and PCI can often be outpatient. So the general idea of “disruption by worse, but cheaper” holds for this example.

Interestingly, more recent data to 2016 demonstrated a roughly 40% decrease in both CABG and PCI — possibly due to more effective medicines (like statins and beta-blockers) reducing the need for surgery. So perhaps medical treatment for heart disease is now disrupting PCI, which previously disrupted CABG? Which might shift revenue back to primary care providers, whether doctors or nurse practitioners or other.

IT-based healthcare disruption

Disruptor of many things

The reason that nowadays we focus on IT when we think about disruption is that for about 40 years, most business disruption in the world has been driven by the IT revolution (further reading: Why Software is Eating the World).

Think about the efficiencies and cost savings made possible by the PC, the internet, cloud computing, and the mobile phone.

Now think about the fact that healthcare has really only adopted one of those innovations: personal computers. Ouch.

Which means that companies that are very, very good at IT are turning their attention to healthcare to see if they can begin picking off the low-hanging fruit from set-in-their-ways old-school health systems. And this includes those using the very latest technology: artificial intelligence (AI).

Some examples:

• efficient lower-overhead systems (like Walmart Health) disrupting hospital-based systems (AND the urgent care companies!) with lower-cost systems that provide only the most commonly-needed services, and that can leverage their enormous portfolio of existing stores for office space. Also remember that Walmart got to be Walmart by being really good at IT (for supply chain management): we can expect efficiencies that most hospital systems can’t approach.

• deep-pocketed tech companies, like Amazon Care, disrupting primary care by using tech to run clinics at lower cost. How can Amazon be successful? By using the same online tools they use as the biggest online retailer in the world and, less well known to the public, as one of the leading providers of cloud computing services to other tech companies, via their AWS division. And they are on a healthcare spending spree lately, having just purchased One Medical [Note: the day after publishing this, Amazon announced they’re shutting down Amazon Care and focusing on One Medical].

• AI-based tools pulling some services away from expensive specialists, with one example being IDx-DR (see next section, below).

These IT-based disruptors are trying to use IT to create simpler, easier to use, cheaper versions of existing healthcare services.

Example of AI disruption: IDx-DR

Digital Diagnostics is a company based in Coralville, Iowa, far from tech centers like Silicon Valley or Boston. Using artificial intelligence computer vision techniques, they’ve built IDx-DR [update: since renamed LumineticsCore]: the “first and only FDA authorized AI system for the autonomous detection of diabetic retinopathy.”

Translated, that means they have a machine that can take pictures of your retina and determine if you have diabetic retinopathy (DR, a debilitating problem that often occurs in poorly controlled diabetes).

Before IDx-DR, the only way to diagnose DR was for a very highly-trained and highly-paid ophthalmologist to look at your eyes. With Idx-DR, the diagnosis is automated and requires no input from any doctor, much less an expensive specialist.

This is the first machine capable of making a medical diagnosis by itself, and according to the website is “now part of the American Diabetes Association’s Standard of Diabetes care”.

Interestingly, it is marketed as a tool for primary care clinics, who can now avoid sending diabetic patients to expensive ophthalmologists for this diagnosis. That is a boon to primary care docs (and their patients) — but with 37 million diabetics in the US alone there is no real technological reason why this shouldn’t eventually find its way to a drugstore near you, right next to that automated blood pressure cuff.

Example of AI disruption: ChestLink

Another great example of AI-based disruptive innovation has been developed by Oxipit. Their ChestLink app, which is approved by European regulators but not yet the FDA, uses AI to read chest x-rays. It is apparently sensitive enough to identify the presence of problem, but not (yet) as good as a human radiologist to identify what the specific problem is.

ChestLink’s main use, as a result, is to automate the reading of the up-to-80% of chest x-rays that are normal, freeing up radiologists to focus on those images flagged by the software as suspicious. Note again that the software isn’t necessarily as good as a radiologist in identifying a specific problem, but it is good enough for a very common task that takes up a lot of human radiologist time.

How is this disruptive? Well, because using the software to verify x-rays as normal is so much cheaper than using a radiologist for the same purpose:

1. rich countries with radiologists won’t need as many, so the cost of this one small facet of healthcare (i.e. reading chest x-rays) will go down.

2. in poor countries, the software isn’t competing against radiologists, it is — in Christensen’s terms — competing against “non-consumption.” It can thus create a new market for the use of chest x-rays, bringing that beneficial medical technology to many more people.

From specialists to generalists, from physicians to nurse practitioners, from hospitals to supermarkets

In Christensen’s 2000 article (echoed later in his 2016 book The Innovator’s Prescription) Christensen presciently wrote about many of the now-established trends in healthcare:

We need diagnostic and therapeutic advances that allow nurse practitioners to treat diseases that used to require a physician’s care, for example, or primary care physicians to treat conditions that used to require specialists. Similarly, we need innovations that enable procedures to be done in less expensive, more convenient settings—for doctors to provide services in their offices that used to be done during a hospital stay, for example.

All of this, of course, is commonplace in the healthcare of 2022: urgent care centers staffed by nurse practitioners, outpatient surgicenters, more techs working throughout healthcare.

But Christensen wrote that in the world of 2000: Google was 2 years old, AI was still frozen in a decades-old “winter”, and the iPhone existed only in the mind of Steve Jobs. The advent of two decades of consumer technology has not only accelerated the healthcare changes that Christensen envisioned, but also accelerated something he doesn’t seem to have thought much about: selfcare. That is, the ability of the consumer to treat their own illness, and to maintain their own health.

And in my next post, I’ll talk about healthcare moving not from doctors to nurse practitioners, but from nurse practitioners to . . . us.

Stay tuned.

This post draws on themes from my keynote talks.

Healthcare in the United States is a $1.27 trillion dollar industry known for

1. very high prices

2. terrible customer service (long waits, short visits, poor coordination)

3. impossible-to-understand billing

4. antiquated information technology (e.g. using fax machines in 2022)

While the saving grace for a long time was that the healthcare system got outstanding results, that doesn’t seem to be the case anymore. While the US spends much more on healthcare than other rich countries, for example, the country is number 46 for life expectancy at birth — right behind Estonia.

And a recent study published in the Journal of the American Medical Association demonstrated that for many common health issues (e.g. colon cancer, heart attack, infant mortality), even the richest Americans often had worse healthcare outcomes than the average citizen of many European countries.

Ripe for disruption — whatever that is

Whenever you have a very expensive product with major problems, there is always an opportunity to find ways to improve or replace it. One important type of replacement using technology is called “disruption”, and it is one of the main strategies that Big Tech companies and others use to take over major segments of whole industries (see: advertising, retail, music, travel, photography, etc).

Unfortunately, most doctors, nurses, PAs, techs, and others aren’t familiar with the term — and are unaware of how technological trends have already begun disrupting their current business models.

The purpose of this and some following posts is to give folks working in healthcare a framework to better see what’s happening in healthcare and better prepare for the future — a future which will be very different from the traditional doctor-and-hospital-centric healthcare model.

What does “disruption” mean?

If the fire alarm goes off at the grocery store while I’m looking for tomatoes, that would definitely “disrupt” my shopping — but this is not the disruption I’m talking about. Likewise, I don’t just mean the introduction of something “new” or “innovative.” And I definitely don’t mean a new style of denim shorts:

I’m using the concept of disruption as defined by the late Clayton Christensen in his incredibly influential book The Innovator’s Dilemma (TID), originally published in 1997. TID talks about how technological change can sometimes improve existing companies’ products, while other times putting such companies out of business.

As a way to explain these two very different outcomes Christensen divides tech innovation into two categories: sustaining and disruptive.

Sustaining innovations

A sustaining innovation is a new technology that helps existing companies (“incumbents”) make their products better for their existing customers.

LED light bulbs are a great example of a sustaining innovation. They’re definitely way better than incandescent bulbs — much longer-lasting, way more energy efficient, cooler — and they’ve helped big incumbent light bulb manufacturers like Philips and GE to make better products, and serve their customers better, and to increase their profits. This sustaining innovation makes incumbents more profitable.

Other examples of sustaining innovation are not hard to find: many of the products we use everyday get better over time. Detergent pods? New versions of your phone’s operating system? Bigger (flatter!) TV screens? Better cushioning material in your running shoes? Self-checkout at the grocery store? Waterproof/breathable rain gear?

None of these innovations are putting anyone out of business. They’re “just” incremental improvements that make useful products even better.

Disruptive innovations

A disruptive innovation is a new technology that lets new companies make a product or service more affordable and accessible to more people in a way that puts incumbents out of business.

Characteristics of disruptive innovations usually include:

1. they are typically worse in many respects than existing tech (this one is a bit counterintuitive, but read on)

2. they are typically better in at least one respect, usually some combination of price and simplicity

3. by being cheaper, or simpler to use, they create a new market of users that couldn’t afford to participate using older technologies

4. they are ignored — or laughed at — by incumbents because they don’t (initially) make much money. Plus, they’re clearly worse (see #1, above).

A great example of disruptive innovation that most of us have lived through is Netflix’ original business model (before they became a streaming giant). You might remember that when Netflix came online in 1997 there was no streaming (the internet wasn’t fast enough yet): Netflix was a DVD rental company. They pioneered a new way to rent DVDs, by ordering them online and receiving them through the mail, which was in direct competition with Blockbuster and other physical video rental shops.

How was this classic disruption? Let’s go through the points:

worse — the first thing I thought when Netflix came out was “who wants to wait 3 days to get a DVD in the mail??” Most people liked the instant gratification of going to the video store and walking out with a movie a short time later. Netflix was definitely much worse in this respect, and was the butt of many jokes when it started.

better — the thing people hated most about Blockbuster and other video retail shops was . . . the late fees. You’d get the DVD or VHS tape for cheap but if you forgot to return it on time you were nailed with a big late fee. In fact, according to Netflix lore, Netflix founder Reed Hastings got the idea when he forgot to return a copy of Apollo 13 on time and got hit with a $40 late fee. Netflix had NO late fees: you could keep a DVD as long as you wanted and just paid the same subscription fee each month.

new market – the market for Blockbuster was obviously limited to places with a Blockbuster store. Netflix, on the other hand, could bring in more customers because they were (1) cheaper and (2) available to anyone in the US that could get mail (i.e. everyone).

ignored/laughed at — a lot has been written about why Blockbuster didn’t/couldn’t adapt to the challenge from Netflix. Christensen would say that it’s because Netflix was offering a mail-order subscription service that — at least initially — Blockbuster’s customers weren’t asking for, and Netflix wasn’t making much money doing it. In fact, in 2000, the Netflix founders tried to sell Netflix to Blockbuster, and they remember that at that meeting the Blockbuster CEO “was struggling not to laugh.”

Netflix initially got subscribers who were angry about late fees, or who didn’t have a good video shop nearby, but as they got more popular their business grew. The last straw was in 2007, ten years after Netflix started, when they announced that they would start streaming movies.

This removed the last objection anyone had to Netflix (having to wait for the mail), and actually made them a faster option than Blockbuster. Three years later, as everyone got used to picking and immediately watching movies without getting up from the couch, Blockbuster went out of business.

Other examples of disruptive innovation include social media — which has put many traditional news media like newspapers out of business — and electronic messaging like email, WhatsApp, and Skype, which has nearly eliminated traditional paper-based messaging. Not to mention the smartphone — which by making computing simpler and cheaper (especially with low-cost Android phones) has brought BILLIONS of new computing users into the market.

Sometimes people argue about whether a new technology is sustaining or disruptive, and the answer can depend on who, exactly, is getting disrupted. The iPhone, Tesla, Uber . . . all of these clear innovations have had people arguing both sides. To read more about the example of Tesla, try “Is Tesla Disruptive” by Ben Evans. To read why even Clayton Christensen, the father of disruption theory, got the iPhone’s disruptive potential wrong, try “What Clayton Christensen Got Wrong” by Ben Thompson.

So what does this have to do with health?

When you work in a very expensive system with arguably poor results, it’s a natural question to think “how can we improve this?”

If you work outside the system, and you understand disruption theory, the natural question is “how can we disrupt this system?”

These questions are being asked throughout healthcare, and by companies outside healthcare. We’re seeing sustaining innovations like better drugs and electronic medical records, as well as potentially disruptive innovations like urgent care centers run by Walmart, and smartphone apps that treat depression without a human therapist

In my next post I talk about the many, many ways that information technology is acting as a catalyst for both sustaining AND disruptive innovations, and the implications for traditional healthcare.

This post draws on themes from my keynote talks.

Picking up the pace of approvals

Not too surprisingly considering the hype around AI — and its potential —  we can see that the pace of FDA approvals has picked up substantially in the last 10 years. While 2022 is not yet complete as of this writing, if it continues at the current pace it should just exceed 2021:

FDA AI-Powered Medical Device Approvals by Year

Lonely at the top

With a bit more digging, we can also see that the number of companies receiving at least one approval is on the upswing, as well:

Number of companies receiving at least one FDA approval by year

Despite the increase in the number of companies playing in the space, success in FDA medical AI approvals also seems to be becoming more concentrated.

For example, in 2022 so far, just 10 of the 55 companies winning at least one approval (18%) were responsible for more than half of total approvals. Five years ago, that percentage was 55%:

Percent of Companies Winning Approval Each Year Getting 50% of Approvals

Your grandfather’s medical AI?

The most surprising point hidden in the dataset was that a lot of the action in medical AI isn’t coming from Big Tech companies known for AI like Apple or Amazon or Google/Verily but from much much older Big Med Tech industrial companies, with GE (founded 1892) and Siemens (founded 1847) taking the number 1 and 2 spots, respectively.

[Note that I’ve counted related company divisions as a single unit. For example: GE Medical Systems, GE Healthcare Japan, and GE Healthcare are just counted as “GE”.]

Here are the companies taking the top 5 spots for “number of approvals in 2022 so far,” with number of approvals in parentheses (and note that Philips and Canon tied for 4th place with 4 approvals):

Top 6 Companies for FDA Al Approvals in 2022 (so far)

The average age of these top six companies is 89 years old! Leaving out Hyperfine and Aidoc, that figure rises to 130. Not exactly startups.

Nonetheless, the oldsters are being crowded out gradually as the median age of companies getting approvals goes down:

Mean and mean age of companies receiving FDA approval

Sustaining innovation rules medical AI

So we’ve established that some pretty ancient companies are among the dominant players in medical AI. But is the tide shifting towards the startups (i.e. towards disruption)?

I’ll define an “older” company as >= 20 years old.

Certainly we can see that the percentage of companies winning at least one approval each year that are older is decreasing:

Percentage of companies winning at least one approval that are >=20 years old vs younger

This may reflect the increasing simplicity and economy of AI tools available, and the increasing number of coders who understand how to use them: the cost in time and money of playing in the AI space is going down, and that means younger and smaller companies can join the game.

At the same time, older companies gain the benefit of those decreasing costs, too, and they have more resources with which to deploy them. Here’s a chart showing the percentage of approvals going to older vs younger companies:

Percentage of FDA approvals going to companies >= 20 years old vs younger

What the last two graphs shows us is that while younger companies are increasingly in the game, the older companies (with presumably greater resources) are more than holding their own in terms of productivity. They make up a smaller percentage of companies winning at least one approval each year, but are taking an increasing share of those approvals.

So using FDA approvals as the metric, it doesn’t look like there’s much disruptive innovation going on in AI-powered med tech: we don’t see established medical device makers like GE being pushed aside (at least not yet). Instead we see:

1. increasing numbers of new companies entering the space and winning FDA approvals

2. continued dominance by established med tech players as they add AI on to their existing products to improve performance and appeal to their best existing customers.

In other words, classic sustaining innovation.

Does this mean that AI isn’t causing disruption related to healthcare? No, not at all. But, as I’ve explained in previous articles on disruption within healthcare and disruption to healthcare, it shows that the same technology (AI) can be used to

1. sustain current healthcare practices (e.g. GE and Siemens and Canon, along with some youngsters, making hospital-based imaging cheaper, more useful, and more accurate)

2. enable players outside traditional healthcare to shift low-hanging fruit to another part of the healthcare system (e.g. Walmart Health)

3. allow companies completely outside healthcare to shift other low-hanging fruit completely outside healthcare (e.g. ResApp)

Thoughts about the above? Inside knowledge you’d like to share? Feel free to contact me privately.

Further Reading

• If you’re interested in learning more about the FDA approval process for AI medical devices, I strongly recommend “How medical AI devices are evaluated: limitations and recommendations from an analysis of FDA approvals” by Eric Wu et al in Nature Medicine (April 2021).

• A good, but now somewhat out-of-date discussion of the types of AI medical devices, can be found in “The state of artificial intelligence-based FDA-approved medical devices and algorithms: an online database” by Stan Benjamens et al in NPJ Digital Medicine (September 2020).

• Nice discussion of current FDA rules for software-as-a-medical-device (SaaMD) at Emerge (June 2022).

On further investigation, I discovered that the flaws in the maps we are viewing at the Hopkins site, and the NY Times, and the WHO, and many other locations are much, much worse than I understood at first.

Every map I’ve seen obscures the progression of the epidemic, rather than informing. Not just one map, but all of them. Let me explain how.

Epi curves show changes over time; most maps show "right now”

As I pointed out last week, the best method to track changes over time is a plain old column graph: an “epi curve” showing cases per day (optimally case per day of onset of symptoms), like this one at WHO’s dashboard:

You will notice something important immediately: confirmed cases peaked on February 4, almost two weeks ago. This is completely invisible when looking at the Hopkins map.

This is because the Hopkins map only shows data from "right now”, rather than a daily progression of data. If you can only see one frame of the movie, you can’t very well understand the plot, can you?

This is the “chronicity problem.”

The Hopkins map has other issues, as well, including a hyperbolic color scheme that looks like it was pinched from a video game called “CoronaGeddon 2019”, and a choice of scale that leaves giant red circles covering 30% of the country — all of which suggests that the outbreak is more extensive than it actually is (when in reality only 0.002% of people in China have coronavirus).

But let’s focus on the chronicity problem.

Could animated maps make things better?

As I explained last week, I’m not against maps at all in general, and a good map can — obviously — emphasize the geographic element in an outbreak. The chronicity problem is a sticking point, however, so to really “track the outbreak”, you’d need a modern map that allows you to follow changes over time.

Theoretically, those animated maps would let you follow the progression similarly to glancing at the epi curve. And who knows, perhaps adding the geographic element would make them even superior to that plain old column-graph epi curve.

I managed to find a few of these online, from MapBox (a mapmaking software specialist), the London School of Hygiene and Tropical Medicine (an august school of public health and tropical medicine), and HealthMap (a rightly-famed mapping institution from Harvard and Boston Children's). Here are images showing the data from January 20th to February 15th.

Via the London School of Hygiene and Tropical Medicine:

Via MapBox:

Via HealthMap/Harvard:

Aside from the terrible color scheme of the MapBox map (worse even than Hopkins), do you notice something common to all of these maps? Take your time, watch any one of them all the way through until it cycles back to beginning.

Don’t see it? Each of these maps shows the outbreak as always getting worse — even though they are graphing the same WHO data as displayed above in the epi curve.

The epi curve shows that the daily new cases have been dropping for almost two weeks, and yet all three of these graphs do not show that. They show everything, always, getting worse. Every circle in the London School and HealthMapper maps always gets bigger, every color in the MapBox map always gets hotter.

What the heck is going on?

All the animated maps are mapping the wrong variable

At the beginning of this article I showed the epi curve from the WHO dashboard: a graph of daily new cases. But on the WHO dashboard there is another graph, as well: the cumulative graph of daily new cases. Let’s look at the two side by side:

Those two graphs are showing exactly the same data, but the cumulative graph keeps adding new cases to the total of old cases, so it never goes down.

Anyone looking at only the cumulative graph thinks things are getting worse (higher) every day, and could completely miss the nearly-two-week decrease in lab-confirmed cases (stats people reading this will realize that the decrease in new cases is reflected in the leveling out of the slope of the cumulative graph).

So all those animated graphs are graphing the never-decreasing cumulative numbers — leading any reasonable viewer to think we’re doomed — instead of the non-cumulative epi curve numbers.

And so are the non-animated maps

And after a bit more double-checking, I realized that all the non-animated maps suffer from the same problem: from WHO to the NY Times, all of the “track the outbreak” maps only track it cumulatively. Those maps always look worse from one day to the next, and they will until there are no more new cases. Which will likely be months.

I’ll be contacting each of these organizations to see if they can start using non-cumulative numbers instead. We’ll see how that goes. If they change them, the animated maps actually would be a useful way to view the progression of the outbreak.

Meanwhile, stop for a moment and think about how much needless alarm has been triggered by the ever-worsening maps.

Important Note: with the SARS epidemic (which was a related virus), the initial peak in China was later exceeded by the worldwide peak — likely because there were many more people in the world than in China who could possible get infected. Although we have no indication that that will happen with COVID-19, it’s certainly possible. By pointing out the decline in cases to date, I’m not saying “it’s all over” — but I know that IF it does peak again outside of China we’ll need to use the right visualizations to respond properly.

Update Feb 19 — Today the London School quietly changed their maps to show DAILY (not cumulative) cases. That means their maps look better as the outbreak wanes.

If you’ve been following the outbreak, you have probably noticed the predominance of mapping in data visualization about the virus. Maps are great, of course, but I suspect that the huge explosion in GPS/GIS/mapping capabilities over the last decade has given us all a “hammer in search of a nail”, applying maps even when better dataviz tools might apply.

How about this much-shared map from Johns Hopkins Center for Systems Science and Engineering (CSSE), which has been promoted by CNET, ZDNet, and the NY Times, among many others :

Or these from CNN and the NY Times:

Although these maps are promoted to help readers “track the outbreak”, they all illustrate why maps are not the primary method used by epidemiologists for tracking outbreaks:

Maps show a single point in time, not a progression

You can’t look at a snapshot and know what happened before that snapshot. Likewise, a single map just shows you “now”. To follow progression using maps you need to view multiple copies of the same map over time and ascertain the changes. The NYT/Hopkins maps, however, do not allow you to view changes over time.

Undaunted, I took to the Internet Archive to get the NYT map for the last six days — which required 30 minutes of compiling, cutting, and pasting. Here’s what I found:

The New York Times tells us that we can use this map to “track the spread of the outbreak”. Can we? I challenge you to look at those six days of maps and tell me what has changed in this outbreak.

But the outbreak has changed, and we’ll see that below.

Maps distort the data

As noted above, the NYT/Hopkins maps make it look like the entire country of China is engulfed by nCoV, but there are only about 35k cases in a population of 1.4 billion. Only 0.002% of people in China have coronavirus!

Likewise the CNN map of “coronavirus’ global spread”, above, seems to show that North America, Europe, Australia, and Russia are, along with China, solidly infected — when there are a total of 14 cases, for example, in the US, out of a population of 330 million.

Sequential maps might be a good way of showing the international expansion of the disease IF it was spreading substantially on an international basis but it isn’t (at least not yet): there are fewer than 400 cases of this coronavirus outside of China. In the US, for example, there are 14. One week ago there were 7. Not exactly a wildfire.

What’s better than a map? An epi curve

What makes the overwhelming reliance on maps so strange is that we have a tried and tested tool to compare values over time. It’s called a column chart — or in the context of an epidemic, an “epi curve.” According to CDC:

An epi curve is a visual display of the onset of illness among cases associated with an outbreak.

And it turns out that we do have an epi curve for novel coronavirus, from the hardworking public health pros at the World Health Organization (technically, since this is "date of report" it's not a true epi curve, but it still works for this example). In the epi curve below, I’ve circled the same six days depicted in the six sequential maps above.

What? Hey? Wait a minute: the number of cases is trending down?!

Yes, the number of reported cases has dropped for 3 of the last 6 days [note: this trend has continued, as of February 13]. There could be many explanations for this, but the most likely one (especially at a stage when the response in China has now had time to gin up testing capabilities and hospital beds, etc) is the simplest: that cases are dropping. And that information appears nowhere at the CSSE dashboard, or at the NY Times coronavirus page. Or at any of the other major news sites that I’ve seen.

WHO has another epi curve, somewhat buried in its daily “sitrep” (situation report) showing the cases outside China:

And you will note that this epi curve is ALSO going down.

If maps are so bad at tracking outbreaks, why is the news full of them?

Of course, I’m not against maps. I’m a map nerd from way back. But to include maps and exclude epi curves leaves readers and viewers dangerously under-informed.

We can see above that

1. Maps can exaggerate the extent of the coronavirus outbreak: China has a lot of cases, but it’s a tiny fraction of the population — but the most-shared maps make it look like everyone in the country is infected.

2. The most useful dataviz tool — the epi curve — is missing in action, and not shown on any of those media sites.

3. Very likely the most important fact about the novel coronavirus outbreak — that it may have peaked — is completely obscured by the data visualization tool selected by every single major media outlet.

I don’t know why the media is making these maps the center of coronavirus data visualization, but if I had to guess I would say it’s because the last ten years have given us a LOT of legitimately cool map technology and maps look a lot cooler than column charts (especially in dramatic black and red).

And as the old journalism saying goes, “if it bleeds it leads” — and maps make things look worse.

But including maps with no epi curve distorts and hides crucial info about this important story, which should be exactly the kind of public disservice that an outfit like the New York Times wants to avoid.

Read more about the problems with coronavirus and data visualization in the followup piece:  CoronaGeddon 2019: Why Every Map You've Seen of the Outbreak is Wrong