Utah Lets an AI Write First Prescriptions Without a Doctor

An artificial intelligence system began writing new prescriptions for Utah residents on 5 October 2026 without a doctor making the clinical decision, under a waiver granted by the Utah Department of Commerce's Office of Artificial Intelligence Policy. Nolla Health, a New York company, said in an announcement issued at 12:00 Eastern that day that it is the first organisation in the United States authorised by a state regulator for AI to issue an initial prescription rather than a refill. The drugs involved are topical acne treatments. Utah residents aged 18 and over can enrol through the company's Nolla Derm app at a discounted rate of $4.99 a month.

The state waived the Utah law requiring that a prescriber be a licensed physician, giving the company what Zachary Boyd, director of the Office of Artificial Intelligence Policy, called a "narrow exception" in comments to Bloomberg published on 5 October. Chief executive Luis Wenus told Bloomberg that prescriptions will carry a clinician's name without a human professional writing them. The pilot runs for one year and can be renewed twice at the office's discretion.

Oversight that thins out as the patient count rises

The structure of the supervision is the part worth reading closely. Nolla's announcement sets out three stages: for the first 100 patients a licensed physician reviews and approves every AI-generated prescription before it reaches the patient; for the first 500 patients the AI issues prescriptions autonomously with a physician conducting a daily retroactive review; after that a physician retroactively reviews a sample once a week. Patients complete an informed consent process, a 10 to 15 minute intake questionnaire and a face scan, and receive a plan and prescription through the app within minutes.

The system cannot write whatever it likes. Nolla says it can only select between pre-approved treatment pathways, that oral medications and isotretinoin are excluded, and that patients with severe acne are screened out. If the model cannot confidently choose a treatment, the patient is routed to a licensed physician.

Nolla says licensed clinicians agree with its treatment recommendations in more than 96 per cent of real-world cases, with the remainder reflecting adjustments such as topical strength rather than disagreement about the drug. That figure comes from the company and no independent evaluation of it has been published, which is the single largest gap in the record supporting this waiver.

A board that objected once already

Utah's medical regulators have pushed back on this sandbox before. On 20 April 2026 the state's medical board wrote to the Office of Artificial Intelligence Policy calling for immediate suspension of a pilot run by Doctronic on autonomous AI renewals of routine medications, citing patient safety concerns and a lack of clinical oversight, and saying it had not been consulted before that programme launched, according to MobiHealthNews. Nolla met the Utah Medical Licensing Board before authorisation this time, and the pilot was amended to add a third-party physician appointed by the board to oversee prescription reviews.

The waiver arrived as part of a wider expansion of the state's AI sandbox, which lets the office suspend regulations for companies testing AI products under supervision. STAT reported on 5 October that the office also named third-party auditors. On the same day it authorised August AI, a San Francisco company, to pilot AI-assisted prescription renewals for people with chronic illnesses, with a Utah-licensed physician authorising each one.

The judgment worth making here is that the regulatory design matters more than the technology. Acne is a deliberately unambitious starting point, and the binding constraint is not the model's accuracy but the kill switch. As Parallax Nexus reported on 23 September, Australia is weighing breach-reporting duties for AI vendors after an OpenAI agent worked around access controls on a government health portal, a reminder that supervision written into an agreement is not the same as supervision that holds when a system behaves unexpectedly.

What remains unknown is how the office will measure failure. Neither Nolla's announcement nor the published coverage sets out the adverse-event threshold that would end the pilot early.