Nolla Health: Utah Lets AI Prescribe Acne Creams

Nolla Health: Utah Lets AI Prescribe Acne Creams

Utah has become the first U.S. state to let an artificial intelligence startup make prescribing decisions with only limited supervision from human doctors. The company is Nolla Health, and the first condition it will treat is acne.

Nolla announced the approval in a blog post on October 5. Founder Luis Wenus went further on X. He described it as the first regulatory approval in the U.S. "(and possibly the world)" for an AI to issue initial prescriptions. "This makes Nolla the first ever actual end-to-end AI doctor," he wrote.

How the app works

The service runs through Nolla Derm, an iOS app. New users complete an intake process before they can use it. Once approved, they describe their acne to an AI agent and upload a facial scan that shows how severe it is. The model then chooses a treatment from a shortlist of topical medicines that physicians have already approved.

Patients can collect the prescription at a local pharmacy or have it shipped home. They take a daily skin scan so the app can track progress, and the treatment plan is adjusted monthly based on how well it is working. Any request outside the system's narrow limits goes to a licensed human physician.

"Until now, a licensed clinician signed every single prescription before it went out," Wenus said in a video introducing the app. "Our AI can now make that prescribing decision itself."

Tight limits on what the AI can prescribe

Four bodies signed off on the pilot:

  • the Utah Office of Artificial Intelligence Policy
  • the Utah Division of Professional Licensing, the state agency that licenses professionals, including doctors
  • the Utah Medical Association
  • the American Academy of Dermatology

The scope is small. The AI may only prescribe creams for topical acne. Isotretinoin and all oral medications are excluded.

The rollout has three stages:

  • Stage one: Two physicians approve every AI-generated prescription before the patient receives it. This continues for at least 100 patients.
  • Stage two: If no serious problems arise, the AI issues prescriptions directly, and doctors review them afterwards. This phase covers 500 patients.
  • Stage three: Physicians review at least 10% of automated prescriptions each month.

The pilot is part of Utah's AI Learning Lab. This sandbox program gives AI companies relief from certain state laws so they can test products under supervision. The state is careful to say what that does not mean. On its authorized-pilots page, the Office of Artificial Intelligence Policy says that taking part "does not mean the State of Utah has approved, endorsed, certified, or vouched for the product."

Acne first, then what?

Zaid Fadul, Nolla's chief medical adviser, told Inc.com that acne is only a starting point. If the pilot succeeds, the company wants to expand to other common conditions. He named three criteria: doctors understand them well, they can be assessed remotely, and the drugs used to treat them have strong safety profiles.

"Beyond dermatology, I think conditions like the flu and uncomplicated urinary tract infections are worth exploring, particularly because testing can provide objective information that can be considered alongside a patient's symptoms and medical history," Fadul said. "The key is having very clear guardrails."

Within ten years, he expects AI to handle much of routine care, including diagnostics, treatments, prescription refills and routine testing. He does not expect it to take over serious conditions, life-changing diagnoses, major procedures or high-risk medications.

He puts stimulants, opioids and benzodiazepines in that last group. One reason is that some patients might try to manipulate the models that assess them. "Frankly, it may take a long time before we have enough evidence and the appropriate safeguards to support autonomous prescribing of some of these drugs," Fadul said.

Our Take

The design of the pilot matters more than the "AI doctor" label. Utah has not handed over open-ended medical judgment. The model chooses from a short list vetted by physicians, for one low-risk condition. Human review is reduced only after the pilot reaches fixed patient numbers. This follows the pattern of safer agent deployments elsewhere: narrow permissions, hard limits set outside the model, and escalation to a human for anything unusual. It also fits the wider move toward vertical AI built for a single domain.

Fadul's point about manipulation deserves attention. Patients have a clear incentive to game a prescribing bot, and work on stress-testing chatbots shows how hard misuse is to anticipate. Public mood is a factor too, given polling that suggests most Americans want AI to slow down.

There are three things to watch. The first is whether Nolla reaches stage two without serious incidents. The second is whether the company or the state publishes the results. The third is whether other states adopt Utah's sandbox approach for healthcare AI.