Utah lets AI prescribe acne treatment without a doctor: inside the first real health sandbox in the US

By Zendoric · October 10, 2026.

On Monday, October 5, Utah expanded its regulatory sandbox for healthcare artificial intelligence —a space where the state temporarily exempts certain technologies from medical licensing requirements so they can be tested on real patients, under supervision— with five new agreements and a network of independent evaluators tasked with assessing their safety and efficacy, according to the state's official announcement reported by Newsweek. Among the pilots, two stand out for what they imply: Nolla Health could end up prescribing certain acne treatments without a doctor's approval, and August AI will test automatic prescription renewals for chronic illnesses, though in this case with mandatory medical authorization.

Intermountain Health and University of Utah Health, two of the state's largest health systems, also signed "umbrella" agreements to speed up the review of future pilots. Intermountain said at Monday's press conference that it already has more than 300 AI initiatives underway, a figure that gives the real measure of adoption even before this announcement. The move falls under Utah's Office of AI Policy and its Pro-Human AI initiative, and the state's stated expectation is that the evidence generated will serve to draft permanent regulation and, potentially, a model for other states.

The evaluation is being handled by six independent entities; one of them is the Coalition for Health AI (CHAI), an organization that has spent years developing voluntary standards —its well-known "nutrition labels" for medical models. Its chief executive, Brian Anderson, and its general counsel, Brenton Hill, explained to Newsweek that CHAI's role is to convene evaluators with the necessary technical capacity, not to certify final approvals: the decision and the ultimate responsibility remain with the state. "We're not going to roll all of this out at once. We're going to test it one by one to make sure it's safe and effective," Hill summed up.

What strikes us as relevant about this episode is the method, not just the headline. As we have noted before at Zendoric, much of the real brake on AI in the United States is arriving through indirect channels —consumer protection laws, competition law, even rulings that apply doctrines far older than AI itself— rather than through AI-specific legislation. Utah is doing exactly the opposite: it is building regulatory infrastructure ex ante, designed to generate real data before legislating firmly, case by case. It is a healthier approach than the usual either/or between banning and letting it run, and it is exactly the kind of evidence-based governance we advocate: neither regulatory panic nor a free-for-all.

That said, it should not be idealized. Hill himself admits that something as basic as the legal liability structure if one of these systems fails is still unresolved: who answers among the various parties involved remains an unknown that shapes how much organizations dare to deploy. Nor is it clear how patients' actual trust will be measured: asked about it, Hill replied that he does not have that information on whether a formal mechanism will exist, and Anderson merely expressed his hope that patients' "sentiment" and experience will be collected. And the Nolla Health case —delegating the prescription of a drug without medical review, even for a low-risk treatment such as acne— is a real precedent for clinical AI autonomy, not a theoretical exercise: none of these uses would be legal in Utah without the sandbox exemption, as Hill himself notes.

The argument that interests us most in this story is the one about access. Anderson frames it as the way to bring care to patients in rural areas of Utah who today would have to travel hundreds of miles to see a healthcare professional, and he mentions, as an example of the kind of models that would need evaluating, an algorithm that identifies patients at high risk of sepsis. It is the kind of case that connects with Zendoric's underlying thesis: well-governed healthcare AI does not replace the doctor, it redistributes care capacity that is scarce today. If the evidence Utah generates shows that certain tasks —a prescription renewal, a risk screening— can be delegated safely, the result is not "fewer doctors" but more doctor-hours available for what genuinely requires human judgment: complex cases and the relationship with the patient. That is, in miniature, the kind of healthcare abundance we believe AI can deliver in the long run, even if the road there runs through uncomfortable pilots like Nolla Health's.

Utah is also exploring reciprocity between states: if another state opens its own sandbox, a technology already approved and in use in Utah could be recognized without repeating the entire process; according to Hill, there are conversations underway about that possibility. If that takes hold, what is today a local experiment becomes the de facto regulatory mold for healthcare AI in the US, in the absence of a federal law arriving in time. The value of this episode lies not so much in whether Nolla Health will end up prescribing for acne without medical supervision, but in the fact that, for the first time, someone is generating the data needed to answer with evidence —and not with ideology— the question that really matters: how much can we trust these systems, and under what conditions.