This analysis synthesizes 8 sources published the week ending Oct 13, 2026. Editorial analysis by the PhysEmp Editorial Team.
Utah is now writing into contracts how much physician attention an AI prescription needs, and the amount shrinks on a schedule. Under the terms Nolla Health described for its acne pilot, a licensed physician approves every AI-generated prescription for the first 100 patients. Through patient 500, the software prescribes on its own and a physician reviews its decisions daily, after the fact. After that, a physician reviews a weekly sample, according to Chief Healthcare Executive.
Most coverage ran with the headline that a state let AI prescribe without doctors, but doctors are still in these pilots. What changes is the work they do, how much of it there is, and who carries the risk when a decision they never saw goes wrong. For anyone following AI in Physician Employment & Clinical Practice, Utah’s expansion this month is the first public template for physician work in a model where the clinical decision moves to software and the license stays with a person.
Five agreements, each with a review schedule
On Oct. 5, Utah’s Department of Commerce added a healthcare pillar to its Pro-Human AI Initiative and approved five new agreements, Becker’s Hospital Review reports. Nolla will test initial prescriptions from a short list of topical drugs for adults with mild-to-moderate acne, per Healthcare Dive. August AI will renew routine, non-controlled medications for chronic conditions, starting with 96 drugs and potentially reaching 148 across 76 conditions, at $4 per renewal, Fierce Healthcare reports. Intermountain Health and University of Utah Health received umbrella agreements that speed review of their future pilots.
Each pilot is stage-gated, and the stages are defined by how much physician review falls away. August AI starts with a Utah-licensed clinician reviewing every decision and plans to end with physicians spot-checking 5% of them. Doctronic, the renewal pilot that launched in January, was designed to finish with physicians spot-checking 5% to 10% of cases. None of the three new companies can move to a later phase without written regulatory approval, Healthcare Dive reports.
The number that decides how many physicians these models need is the review rate at steady state, and the agreements already set a target of 5% to 10% of decisions. Accuracy claims matter mostly because they are what moves a pilot down that scale.
What the agreement rates show, and who measured them
The case for lowering review rests largely on company data. Nolla says clinicians agree with its AI’s treatment choice in more than 96% of real-world cases. It also reports treating more than 50,000 patients in Norway in six months while cutting clinician time per patient tenfold. Chief Healthcare Executive points out that these figures came from Nolla’s own release, without independent verification.
Doctronic’s May report is more revealing for anyone thinking about staffing. Its AI recommended renewal in 72% of cases and escalated 28% to a physician, and reviewing physicians agreed with 91% of the renewal recommendations, according to an analysis by the law firm Snell & Wilmer. Physicians judged 69% of the escalations warranted and the rest overly cautious. So more than a quarter of the volume came back to humans, and part of the physician’s job was clearing the software’s false alarms. Those results came from Doctronic’s own physicians; the state has started a separate independent review.
Utah’s answer to self-reporting is a roster of six outside evaluators, including the Coalition for Health AI and Stanford’s Clinical Excellence Research Center, who will stress-test guardrails and inspect decision logs. CHAI general counsel Brenton Hill told Newsweek that “we still don’t fully understand where liability is going to sit,” and said that uncertainty often limits deployment at scale.
Liability follows the name on the prescription
Snell & Wilmer’s reading of Utah law is that sandbox participation does not shift risk to the state, which disclaims responsibility for claims arising from a participant’s involvement. While a physician reviews each decision, that reviewer may serve as the “learned intermediary” product-liability doctrine relies on. The protection is less certain once the AI acts first and a physician looks later. The firm warns that physicians whose names appear on AI-generated prescriptions may face liability for decisions they did not individually make, while vendor terms broadly disclaim responsibility for accuracy and outcomes.
Licensing is unsettled as well. Utah’s corporate practice statute requires medicine to be practiced by a licensed individual, and an AI system cannot hold a license. In April, 11 of the 14 members of the Utah Medical Licensing Board signed a letter calling for “immediate suspension” of the Doctronic pilot. The state declined and formed a joint working group with the board instead.
A physician reviewing a weekly sample is vouching for a process rather than for individual patients. Few employment agreements written today define that role, set a rate for it, or say whose malpractice policy covers the prescriptions the physician never opened.
The pilots require participants to carry specialty malpractice insurance and keep 24-hour incident response plans, Becker’s reports. A physician signing on to review AI output, whether for a vendor panel or a health system building its own tool, needs the same answers Snell & Wilmer lists for employers: who authorizes each decision, whether indemnities address autonomous action, and how credentialing and peer review apply. Recruiters should also note the firm’s warning that a Utah agreement does not travel. Colorado requires human review of certain AI decisions, and Alabama, Iowa and Washington have restricted AI in utilization review.
A physician-supervised model produced different evidence
The same week offered a contrast. In a Lancet study run by Google and Beth Israel Deaconess Medical Center, primary care patients chatted with Google’s AMIE system before urgent visits while a physician watched every exchange, R&D World reports. Among 98 patients who completed both steps, AMIE’s first seven suggestions included the final diagnosis 90% of the time, but its top suggestion matched in only 56% of cases. In 44 cases where physicians read the transcript beforehand, they said it helped them prepare 75% of the time.
That design puts AI in front of the visit and keeps the visit, which is closer to the American Medical Association’s June policy, cited by Becker’s, calling for AI to be integrated into physician-led processes. Federal regulators are testing both paths. KFF notes that more than 200 companies have CMS approval to test AI tools that perform clinical functions, per The New York Times. In looking back at our recent work, we have said too little about patients who consult AI before, or instead of, a physician, and Utah makes that trend harder to set aside.
Doctronic’s agreement runs through October 2026, with a state option to renew for another year, and Becker’s reports the company is in talks to expand to Arizona, Missouri, Texas and Wyoming. Watch the first medication group to clear the 250-prescription threshold and receive approval for retroactive review. That decision will show whether a 91% agreement rate is enough for a state to accept a physician who checks the work afterward, and other states will be looking at the same number when they write their own terms.
Sources
AI prescribing gets a state-supervised test in Utah – Chief Healthcare Executive
Utah’s AI physician experiment expands: What to know – Becker’s Hospital Review
Utah approves slew of health AI pilots for drug management, women’s health – Healthcare Dive
Utah expands healthcare AI sandbox with August AI, Nolla Health pilots, adds third-party evaluators – Fierce Healthcare
When AI Writes the Prescription: Licensure, Liability, and the Corporate Practice of Medicine in Utah’s Autonomous AI Pilot – Snell & Wilmer
What Utah’s AI Experiment Means for Hospitals – Newsweek
Google AMIE study adds evidence of medical AI’s promise with patients, under doctors’ supervision – R&D World
AI in Health Care Moves Toward More Autonomous Roles – KFF