Utah lets an AI write first-time acne prescriptions, a US first

Since October 5, adults in Utah can get a first prescription for acne from an AI in the Nolla Derm app, for $4.99 a month. It is the first time a US regulator allows this. Here is how it works, how doctors stay in the loop and what is still unclear.

For the first time in the United States, a state regulator has agreed to let an AI system write a brand new prescription on its own. Since Monday, October 5, adults in Utah can open the Nolla Derm app, answer a short medical questionnaire, scan their face and, if the software is confident, walk away with a prescription for acne treatment a few minutes later. No waiting room, no video call, no human signature on the first decision once the early test phase is over.

The service comes from Nolla Health, a New York startup, and runs inside Utah's AI regulatory sandbox, the same program that already allowed an AI to renew existing prescriptions earlier this year. The difference this time is big: renewing a medicine a doctor already chose is one thing, choosing the medicine in the first place is another. That is why this small acne pilot is getting far more attention than its scope would suggest.

Card from the djcroman Short: questions plus face scan equals prescription, $4.99 a month, adults in Utah

The three things you need to know

How it works. Utah residents aged 18 and over download Nolla Derm, verify their identity, go through informed consent, fill out an intake questionnaire of about 10 to 15 minutes and take a face scan from several angles. The AI then picks a treatment plan and, if it fits the rules, issues the prescription. The subscription costs $4.99 a month, the medicine itself is paid separately.
Why it is a first. Nolla Health says it is the first organization in the country with regulatory approval for AI to issue initial prescriptions, not just refills. According to Bloomberg Law, the system can currently choose from eight different skin treatments.
The safety net. For the first 100 patients, two Utah licensed physicians approve every AI prescription before it goes out. After that, the AI sends prescriptions directly and doctors review them afterwards, and later only a sample of at least 10 percent each month, plus every case with a side effect or an escalation. Severe acne, pregnancy and other cases outside the pilot are sent to a doctor instead.

What the app actually does

Nolla Derm is not a general chatbot that happens to know some medicine. The company describes a tightly closed system. The prescription flow has no free text box, so a patient cannot talk the model into something unusual. The AI combines structured answers from the questionnaire with its own analysis of the face scan and then selects one option from a short list of topical treatments that physicians approved in advance. It cannot invent a new treatment, change a dose freely or prescribe pills outside the list.

If the system is not confident, it is supposed to stop. In that case the patient is handed over to a licensed physician instead of getting an uncertain recommendation. Nolla also says that whenever a patient messages for support inside the app, they reach a human doctor, not the AI. After the first prescription, the app asks for regular skin scans to track progress, and treatment plans are reviewed and adjusted monthly. Medicine is either delivered to the patient's home or picked up at a pharmacy.

The company backs its pitch with one headline number: licensed clinicians agreed with the AI's treatment recommendation in more than 96 percent of real world cases, and Nolla says the remaining cases were minor changes, such as a different strength of the same cream, rather than real disagreements. That figure comes from the company itself, and it has not yet been checked in an independent, published study.

Why acne, and why Utah

Acne is a deliberate choice. It is extremely common, it is well understood, and the first line treatments are topical creams and gels with decades of evidence behind them. If something goes wrong, the risk is far lower than with heart medication, antibiotics or psychiatric drugs. That makes it a natural place to test whether AI can handle the first step of care, not just the paperwork around it.

The access problem is real as well. In its announcement, Nolla says about 100,000 people in Utah have active acne at any given time, but only around one in ten ever sees a dermatologist. The company puts the average wait for an appointment in Utah at 61 days, says 11 counties have no dermatologist at all and that a single visit can cost up to $300. Untreated acne can leave permanent scars, so a long wait is not just an inconvenience.

Illustration: a long queue outside a small dermatology clinic and a wall calendar full of crossed out days, while one person gets treatment through a phone app

Utah is also the state that built a legal path for this kind of experiment. Its Office of Artificial Intelligence Policy runs a regulatory sandbox that can temporarily relax certain rules for companies that agree to strict conditions, reporting and the possibility of being shut down. Nolla says it worked with that office for months and also met the Utah Medical Board before launch, so the board's feedback could shape the final program. Zach Boyd, who leads the AI policy office, called the pilot a significant step for access to care while keeping oversight and patient safety in focus.

The step by step safety net

The most interesting part of the program is not the AI model, but the way human control is reduced in stages. According to the published details, the oversight works like this:

Stage one: for at least the first 100 patients, two Utah licensed physicians review and approve each prescription before it is sent.
Stage two: for up to around 500 patients and at least eight weeks, the AI submits prescriptions directly, and physicians review the cases afterwards.
Stage three: physicians check a retrospective sample of at least 10 percent of the volume each month, plus every case that involved a side effect or an escalation.

Dr. Zaid Fadul, Nolla's chief medical advisor, framed the whole design around control. In his words, the real test for AI in medicine is not how smart it is, but whether someone other than the company can switch it off. In this pilot, both the state and the supervising physicians can.

Illustration: three steps, from two doctors watching every AI prescription, to one doctor reviewing finished prescriptions, to a doctor checking a single sample

Utah has been here before, and doctors pushed back

This is not Utah's first AI prescribing experiment. Earlier in 2026, the state partnered with the startup Doctronic on a pilot that lets an AI handle routine renewals of medicines that a human doctor had already prescribed. That project became controversial fast. In April, 11 of the 14 physicians on the Utah Medical Licensing Board signed a letter asking for the pilot to be suspended immediately. Their main complaint: the board only learned about the program after it was already live, and they argued that even a refill needs a clinical check for side effects, interactions and changes in a patient's health.

The state refused to suspend it, pointing out that in the first phase a physician still reviewed every refill decision. Early numbers published by Utah's Department of Commerce showed that the AI recommended a renewal in 72 percent of cases, and reviewing physicians agreed with those recommendations 91 percent of the time. In the other cases the AI escalated to a human, and doctors said some of those escalations were overly cautious.

Nolla seems to have learned from that fight. Involving the medical board before launch, narrowing the scope to a short list of low risk creams and publishing the oversight stages all look like answers to the criticism that hit Doctronic. Bloomberg Law still put it bluntly in its headline: AI can now prescribe acne drugs in Utah, sidestepping doctors. For many physicians, that is exactly the problem.

What we still do not know

Several important questions remain open:

Independent data. The 96 percent agreement figure comes from Nolla. Until outside reviewers look at the cases, we only have the company's word for how well the AI performs.
Who is liable. If an AI prescription causes harm, it is not yet clear how responsibility is split between the company, the supervising doctors and the state that approved the pilot.
Skin tones and image quality. Face scans depend on lighting, cameras and skin types. Any bias in the training data could mean worse results for some patients, and the company has not published a breakdown.
What comes next. Acne is a safe starting point. The real debate begins when companies ask to move to conditions with higher risk or to medicines with more side effects.

Why it matters

This pilot moves the line on what AI is allowed to do in healthcare. Until now, AI in medicine mostly helped humans: it summarized notes, collected information before a telehealth visit or flagged scans for a radiologist. Here, for a narrow group of patients, the AI makes the first treatment decision and a human checks later, if at all.

If the results hold up, other states and other conditions will follow, and the argument for AI as a way to reach people who never see a specialist will get much stronger. If something goes wrong, it could set back the whole idea for years. Either way, the data from this small acne program will be quoted in every future debate about AI doctors.

There is also a cost angle. A subscription of $4.99 a month is a fraction of a specialist visit. For patients without good insurance or without a dermatologist nearby, that price could be the difference between getting treatment and doing nothing.

Dany's take

I like the design more than I expected. A closed list of treatments, no free text, a clear handover to humans and a state that can pull the plug is a much more responsible start than a general chatbot handing out medical advice. Acne is the right test case: common, frustrating, rarely dangerous and badly served by long waiting lists.

What I do not like is that the most important number, the 96 percent agreement, comes from the company itself. Before this model is copied for anything riskier, I want to see independent reviews, data on different skin tones and an honest report on every case that went wrong. Utah has built a real lab for AI medicine. Now it needs to publish the lab results.

And one more thing: the fact that the board of doctors had to fight to even be informed in the Doctronic case shows how fast this is moving. Patients deserve to know when an AI made the call, and doctors deserve a seat at the table before launch, not after.