AI prescribing: what a health brand can say while Utah and the medical boards write the rules

What is being written, and by whom, for ai prescribing#

A fox reads the wind before it crosses open ground. The wind around ai prescribing this week blows from the Federation of State Medical Boards. STAT reported on 14 October that executives from the American Medical Association, the American Hospital Association, Anthropic and the insurer Elevance Health sit on a 22-member AI workgroup convened by the federation. The workgroup's discussions are being kept confidential, and its stated goal is to draft guidelines for local boards, with particular attention to tools that perform clinical functions with limited or no direct physician supervision. The federation's CEO told STAT a draft could appear in weeks to months.

That is the weather system for ai prescribing, and every fox in the thicket should note which way it turns. The workgroup was announced in May, and the announcement itself puts the same phrase in print, stressing limited or no direct physician supervision as the focus. In August the federation's leaders wrote in STAT that AI is not ready to be independently licensed like a physician. Guidelines from a body of boards are not law, but STAT notes they could shape legislation, because states regulate medical licensure individually.

Meanwhile the ground has already shifted in Utah, where ai prescribing left the den and walked into daylight. The state's Office of Artificial Intelligence Policy signed a regulatory agreement for a yearlong pilot with the company Doctronic that began in December 2025. The Association of Health Care Journalists summarises it as allowing 30-, 60-, or 90-day refills for 190 commonly prescribed drugs, with the programme covering refills only. Marketplace reported that the state waived normal licensing rules to allow the AI to make clinical decisions, and that the state's medical licensing board was blindsided and pressed for a pause.

A second Utah pilot, reported by STAT and Runtime Wire on 5 October, covers first-time prescriptions of a short list of topical acne treatments through an app named Nolla. Runtime Wire says the pilot starts with two physicians reviewing every prescription. STAT's wording is that a licensed clinician reviews and signs off, so coverage differs on the number, and you should check the executed agreement before quoting it.

Three registers of claim: prescribe, support, advise#

For a marketing team the useful move is to stop saying AI and start asking which verb the product performs, because ai prescribing is one verb and decision support is another. The word prescribing, supporting and advising sit under three different regulators, and the evidence each one expects is different. Mixing them in one headline is how a clean product gets a hard letter.

Take the middle row first, because ai clinical decision support is where most commercial products live. FDA's clinical decision support guidance, issued in January 2026, sets out four criteria for software that may sit outside device regulation. One of them is that the software is intended for the purpose of enabling such health care professional to independently review the basis for such recommendations. Read it as a product requirement: a black-box score that a clinician cannot interrogate is not the same thing as a transparent recommendation, and the marketing should not blur the two.

The guidance is written around health care professionals. A consumer app that tells a patient what to do falls outside that framing, which is why ai medical advice sits on the third row with the least comfortable legal footing. That is our reading, not FDA's statement, so take it to counsel before you rely on it.

Two other sources matter for copy. FDA's January 2025 draft guidance on AI-enabled device software says it when finalized, will provide recommendations regarding the contents of marketing submissions, which tells you the lifecycle rules were still draft at that notice. And Texas, which runs its own AI sandbox, requires that where AI is used in a health service the provider of the service or treatment shall provide the disclosure. If you sell into Texas, a disclosure line in your copy is not optional decoration.

What the evidence supports for autonomous clinical ai#

Claims need evidence, and the evidence for ai prescribing and autonomous clinical ai is thinner than the pilots suggest. Follow the trail, not the scent of a press release. Start with the review rate, because it changes the meaning of the word autonomous. Marketplace reports that in Utah's refill pilot's first phase doctors reviewed each renewal, and that only 10% of prescriptions will be reviewed by a human in the next phase. The health journalism summary adds that Doctronic is requiring physicians to review the first 250 prescriptions in each drug class before going completely autonomous. Both are secondhand accounts of the agreement, so confirm the final numbers.

Accuracy evidence is mixed in the way early evidence usually is. A peer-reviewed crossover study in Cell Reports Medicine had pharmacists review 91 prescribing-error scenarios with and without model support, and its authors concluded that the need for further validation before integrating such tools into routine clinical practice remains. A Frontiers in Pharmacology study of complex intensive-care regimes was blunter: using present models as a clinical support tool warrants caution.

On the patient side, a preprint posted in September compared six chatbot set-ups with an Australian government symptom checker across 45 vignettes. It reports that the checker achieved 48.9% accuracy against 73.3% to 86.7% for the chatbots. Read the caveat before you reach for the headline: the authors say the symptom checker's conservative defaults may have disadvantaged Healthdirect, and preprints have not been through peer review.

Taken together the studies point one way for the copywriter writing about ai prescribing: strong results on a vignette set are an encouraging sign, not substantiation for a patient-facing claim. Vignettes are written by clinicians for tests; real patients arrive with missing information and odd phrasing. Say what was tested, how many cases, who graded it, and what the authors say is missing.

Four rules for health AI copy until the guidelines land#

None of the rules below for ai prescribing and its neighbours requires waiting for a regulator, and each one keeps a prowling journalist at bay. They are folkfox's recommendation for AI health teams, built on the sources above, and they are deliberately boring because boring survives a hard letter.

Rule three deserves a pause, because ai prescribing pilots show the human is a dial rather than a switch. Autonomous is a word that carries a claim about the absence of a human, and the pilots show that absence is a dial, not a switch. If your copy says fully autonomous, and the brush of a regulator's pen finds that your agreement says physicians review the first 250 prescriptions per drug class, a regulator or a journalist will read the gap as the story.

We have covered the claim-ladder logic before in the ambient scribe claim ladder and the device side in AI medical devices and the claim a regulator can quote back at you. For the UK angle on what advertisers can no longer say, read what the NMC just wrote down. The sector view is on the healthcare marketing page, and a conversation starts at the contact page.

Frequently asked questions#