Summary
A Utah pilot of Nolla Health’s AI acne-prescribing tool is designed to reduce doctors’ review in stages. A BMJ commentary says any reduction in oversight should depend on evidence about patient outcomes, safety and follow-up.
A Utah pilot of Nolla Health’s AI tool for acne treatment is designed to reduce doctors’ prescription review in stages. In a BMJ opinion published on 13 November 2026, five authors argue that clinical oversight should not be reduced until independent, publicly available evidence shows the system provides safe and effective care.
The tool uses a facial scan and a short questionnaire to generate topical acne treatment plans and prescriptions for eligible patients. The commentary says the pilot, launched on 5 October, excludes severe acne and oral medicines, including isotretinoin. Its focus is therefore a limited group of acne treatments, not autonomous prescribing across medicine.
Doctor review is planned to change in stages
During the pilot’s first stage, two doctors are to approve every prescription before it is issued. If the pilot advances with state approval, doctors in the second stage will review every prescription after it has been issued. In a possible final stage, they would review at least 10% of prescriptions each month, as well as every case involving an adverse event or escalation.
That sequence changes when a clinician might catch a problem. Review before prescribing can identify an error before treatment begins; review afterwards may detect it only once care is under way. The BMJ authors argue that each step away from prior approval should be supported by evidence.
The commentary reports that Nolla Health says clinicians agreed with about 96% of 1,000 evaluated treatment recommendations. The remaining disagreements involved minor changes to drug strength. The authors caution that agreement alone is difficult to interpret: different treatments can be clinically appropriate for the same acne case. Evaluation therefore needs to distinguish reasonable alternatives from decisions that could harm patients.
Measure patient outcomes, not just recommendation agreement
The authors say assessments should track whether acne improves, whether adverse effects occur, and whether people who need further care are identified and complete referrals. They also call for accounting for patients who stop using the service or are lost to follow-up, since results based only on people who remain engaged may miss unresolved health problems.
Performance should also be reported across patient groups, including by skin tone and image quality, with the number of patients in each group and uncertainty around the estimates. These details matter for an image-based tool: an overall score can conceal differences in how well it works for particular groups.
The commentary says Utah requires quarterly public reports and has recruited independent evaluators to audit AI pilots, while their evaluation criteria were still being finalised. The authors argue that findings should be publicly available and independently assessed before clinician review is reduced. They also call for reassessment if the AI model or clinical protocols change.
The central issue is not whether an AI system can produce a prescription recommendation, but whether its use improves care without creating avoidable risks. For this pilot, the proposed move from approval of every prescription to selective review makes the quality and transparency of patient-outcome evidence especially consequential.