Utah's AI-prescribing pilot puts the state at the center of a

A first-mover bet with nationwide consequences

Utah is about to become the first US state where an artificial intelligence system can look at a patient, make a diagnosis, and send a prescription to a pharmacy — with no physician staring at the screen in real time.

The program, run by Salt Lake City startup Nolla Health, currently covers only mild-to-moderate acne. That deliberate narrowness is its political armor. Acne doesn’t kill people. But the legal and regulatory precedent being set here will not be easy to contain.

The application costs $4.99 a month. Users scan their faces, complete a medical history form, and an algorithm scores their skin on a severity scale from clear to severe. It then recommends one of eight approved topical treatments. No doctor is involved in the decision — at least not until Phase 3, where a physician will review at least 10 percent of prescriptions each month, retrospectively.

That word matters: retrospectively.

The phased rollout reveals the real risk

Nolla Health CEO Luis Wenus has framed the pilot as a cautious experiment. The first 100 prescriptions go through a human doctor before reaching the pharmacy. The next 400 patients are reviewed weekly by a physician after the fact. Eventually, oversight drops to a monthly 10 percent sample.

But the direction of travel is unmistakable. The model trends toward full automation, with human involvement shrinking from gatekeeper to auditor. That is the architecture the company is building toward — and Utah regulators are approving it.

This structure also creates a liability fog that existing medical malpractice law was never designed to handle. When a human doctor prescribes a medication, responsibility is straightforward: the physician is accountable under state licensing boards and malpractice insurance. When an AI system makes the call and a human doctor only reviews it after the fact — and not every prescription — who bears responsibility for an adverse outcome? The startup? The software developer? The reviewing physician? The state that licensed the program?

No court has answered that question yet. Utah is about to find out.

FDA oversight hangs in the balance

The US Food and Drug Administration regulates how medications are prescribed and approved, but it has not issued a comprehensive framework for AI-driven clinical decision-making. The agency has granted clearance to some AI-based diagnostic tools, but those devices typically sit alongside human physicians rather than replacing them.

Nolla Health’s model skates along the edge of that distinction. The AI is generating prescriptions, not just diagnostics. And it is doing so in a state-sanctioned program that bypasses the traditional physician-patient relationship.

If the FDA does not intervene or clarify its position, states will fill the vacuum independently. Utah is effectively volunteering as a regulatory laboratory. Other states are watching closely.

There is a competitive dimension here that deserves attention. States that move early on AI healthcare adoption position themselves as destinations for digital health startups. Utah already courts tech companies with favorable tax policy and a business-friendly regulatory environment. This pilot extends that strategy into healthcare — an industry far larger than tech, and one that generates enormous lobbying power.

The acne window is narrow by design

Choosing acne as the entry point is strategically smart. It is a high-volume, low-acuity condition. Most prescriptions are for topical retinoids and antibiotics — medications with side effects limited to skin irritation or dryness, according to Wenus. No systemic drugs. No controlled substances. No emergency decisions.

But precedent rarely stays contained within its original category. Oncology, dermatology, psychiatry, primary care — these are all areas where AI-assisted or AI-led prescribing could be argued as safe and efficient. The infrastructure built for acne treatment is transferable. The regulatory approval process is transferable. The liability questions are transferable.

The first test is simply whether anything goes badly enough to trigger a backlash. If the 1,000-patient pilot runs smoothly with zero serious adverse events, the argument for expansion becomes political rather than scientific.

Errors are already a documented problem

The cautionary cases are real and recent. Last year, a 60-year-old man followed ChatGPT’s advice to replace dietary salt with sodium bromide, developing bromism — a psychiatric condition rarely seen since the 19th century. In March, Google quietly shut down an AI-powered health search feature called What People Suggest after reports that inaccurate medical information was being served to users.

These incidents involved consumer-facing AI tools, not regulated medical programs. But they illustrate a pattern: AI systems hallucinate, they generalize beyond their training data, and they present fabricated certainty. The difference between a bad dietary recommendation and a bad prescription is measured in organ failure and death.

Nolla Health’s safeguards — the phased rollout, the topical-only medications, the eventual retrospective review — are designed to minimize that risk. They are not designed to eliminate it. And they raise the question of whether retrospective review is meaningful oversight at all. A doctor reviewing 10 percent of prescriptions once a month cannot catch systematic errors. They can only catch outliers — and by then, the harm is done.

What other states will do

Idaho, Florida, and Texas have all introduced legislation exploring AI in healthcare. Some bills focus on diagnostic assistance; others push toward autonomous prescribing. None have reached the point Utah has, but several are close.

The first-mover advantage for Utah is real but fragile. If the program produces negative outcomes — even a single serious adverse event — other states will hesitate. If it produces no negative outcomes, the model spreads quickly. Digital health startups will market it. State legislatures will copy it. The FDA will either accelerate its own framework or be forced to respond to a patchwork of state-level rules.

The broader implication is that healthcare regulation in the United States may shift from a federal-to-state coordination model to a race to the bottom, with states competing to attract AI health companies by offering lighter oversight. That dynamic has played out in tech before. It is not guaranteed to play out differently in medicine.

The doctor question is not settled

Physicians are widely considered one of the professions least vulnerable to AI replacement. Automated prescribing in low-acuity conditions challenges that assumption directly. A dermatologist or primary care doctor seeing 20 acne patients a day could theoretically be replaced by a $4.99 app for a significant portion of that caseload.

That does not mean the profession is collapsing. It means a specific slice of clinical work — repetitive, protocol-driven, low-complexity — is becoming economically vulnerable. And the economics are clear: insurers and patients will choose the cheaper option if it is legally available.

Utah’s program is small. Its impact on the medical workforce is negligible today. But the structural argument it establishes — that AI can safely handle certain medical decisions without human oversight — is the foundation everything else will be built on.