Utah pushes health AI deeper into prescribing and women’s care
Utah has widened one of the country’s most closely watched healthcare AI experiments, approving a new set of pilots that could eventually let software help issue prescriptions, renew medications, and deliver virtual pelvic floor therapy with less direct clinician involvement. The latest approvals extend a program built to test how far autonomous systems can move inside clinical workflows before regulators decide whether the safeguards are sufficient for broader use.
The most controversial of the new pilots centers on prescribing. Nolla Health, developer of a dermatology app that evaluates acne from patient photographs, will test whether the system can prescribe initial medications and refills from a short list of topical drugs for adults in Utah with mild-to-moderate acne. At the outset, two physicians will review every prescription before it is sent. Later phases would shift those physicians into after-the-fact review, a sequence that changes the practical meaning of “oversight” from real-time supervision to retrospective auditing.
Acne treatment may appear operationally narrow, but prescription automation in dermatology still raises familiar safety questions: whether image-based algorithms can capture severity accurately, whether they can distinguish acne from other facial eruptions, and whether a limited formulary truly contains enough clinical nuance for adults whose disease burden may fluctuate over time. A topical-only pathway may seem constrained, yet even simple regimens depend on correct diagnosis, dose selection, adherence counseling, and the ability to recognize when an in-person examination is needed instead of another refill.
Utah’s Office of Artificial Intelligence Policy approved the new pilots on Friday. Other programs on the list include one from August AI, which will use an AI system to renew prescriptions for people with chronic conditions, and another from Expect Fitness, which will provide virtual pelvic floor therapy for women with urinary incontinence, pelvic pain, and postpartum pelvic floor problems. None of the pilots has launched yet, and all three will need written approval from regulators before they move into later stages with less human oversight.
A state sandbox built to accelerate testing without full market rollout
Utah created its Artificial Intelligence Learning Laboratory, also called the AI sandbox, through state law in 2024. The state framed the laboratory as a supervised environment where private companies could test and evaluate AI-based tools before wider deployment, while also building consumer protections into the process. Companies enter into agreements with the state that define boundaries, safeguards, and reporting requirements, which makes the program less like a free-form innovation zone and more like a regulated proving ground for clinical automation.
The clinical stakes are clearest in medication management. Prescription renewal appears low risk when compared with diagnosis or procedure selection, but renewal systems can still fail in ways that matter at the bedside: they can repeat outdated therapy, overlook drug interactions, miss a change in renal or hepatic function, or extend a medication that should have been reconsidered after new symptoms emerged. In chronic disease care, continuity is a virtue only when the underlying treatment remains appropriate. Automation can preserve efficiency; it can also preserve error.
The pelvic floor therapy pilot carries a different kind of promise and risk. Virtual therapy for urinary incontinence, pelvic pain, and postpartum pelvic floor problems could expand access for women who face distance, scheduling barriers, or stigma when seeking care. Yet pelvic floor disorders often sit at the intersection of anatomy, neuromuscular function, childbirth history, and symptom severity. Digital delivery may improve reach, but it still depends on accurate triage, patient adherence, and recognition of cases that need hands-on evaluation rather than a remote program.
Physician resistance, liability questions, and the problem of responsibility
The state’s push has not occurred in a vacuum. Earlier AI pilots drew sharp criticism from physicians and other stakeholders, especially after Utah launched a Doctronic program that would renew medications without clinician oversight. The state’s medical licensing board asked that pilot to be suspended, arguing it had not been informed before launch and warning that patients could receive outdated medication without adequate clinical judgment from a human reviewer.
Regulators refused to pause the program, saying clinicians supervise AI systems during the first stage. The tension here is not merely philosophical. It is operational. Who is responsible when a system recommends a medication that is no longer appropriate, or when it renews a drug in the face of a change in symptoms, comorbidity, or safety profile? In a JAMA Health Forum commentary, Stanford law and health policy professor Michelle Mello argued that Doctronic’s contract is unclear about who bears responsibility for AI prescribing errors that harm patients. The liability question becomes unavoidable once software stops acting like a documentation aid and starts acting like a clinical gatekeeper.
Across all three new pilots, the structure of oversight is the central clinical issue. A physician review step before prescribing offers a familiar human checkpoint, but later-stage review after prescriptions have already been sent changes the risk profile substantially. The same is true for renewals. If the system eventually operates with less human involvement, the question is no longer whether AI can accelerate administrative work; it is whether the state can prove that faster decision-making does not outpace safe decision-making.
Those concerns have not slowed Utah’s broader ambitions. Last month, the Office of Artificial Intelligence Policy signed master agreements with University of Utah Health and Intermountain Health. These agreements are not pilots themselves, but they establish the terms under which the state can later approve pilots, signaling that Utah intends to keep expanding the program’s reach. The next phase will test not only the technical performance of the tools, but also whether the state can define a durable standard for oversight before autonomous prescribing becomes routine rather than experimental.
For now, the pilots remain proposals, not active clinical services. Their real significance lies in the boundary they are trying to move: from AI as a support tool toward AI as a participant in therapeutic decision-making. That boundary may hold for a while in mild acne, chronic medication refills, or pelvic floor care. It becomes harder to defend if the same framework is extended to more complex conditions without stronger evidence, clearer liability rules, and a more explicit answer to a basic clinical question: when a machine helps choose treatment, who is actually practicing medicine?