Signals: Utah's AI Prescriber Was More Cautious Than the Doctors Auditing It
Five months into Utah's Doctronic pilot — the first US program authorizing AI to recommend prescription renewals — the state's Office of AI Policy just released outcome data. It is not what AI critics expected.
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.