Utah has opened the nation’s first state-regulated path for an artificial intelligence system to issue an initial prescription without a doctor reviewing the individual case first.
The Utah Office of Artificial Intelligence Policy authorized Nolla Health’s acne-treatment pilot under an agreement signed September 22. Nolla announced on October 5 that Utah residents age 18 and older could begin enrolling in the service.
The first stage still requires two Utah-licensed physicians to review every AI-generated prescription before the system sends it to a pharmacy.
Under the agreement, the AI can begin submitting prescriptions directly only after the pilot runs for at least four weeks, treats at least 100 patients, maintains at least 95% agreement with physicians, misses no required safety stops and records no serious adverse events. Utah regulators must approve the move to the next stage.
The timing remains unsettled.
Utah’s public pilot page still listed the demonstration period as not started on October 6, while Nolla said patients could enroll beginning October 5. The contract defines the commencement date as the date Nolla makes the technology available and gives written notice to the state. That means Utah has authorized a route to autonomous prescribing, but the no-preapproval phase has not clearly taken effect.
What Utah Actually Approved
Nolla’s system does not function as a general-purpose chatbot with open-ended prescribing power. It uses a structured medical intake, five-angle facial images and predefined clinical protocols to assess adults with mild to moderate acne.
The pilot limits treatment to topical acne products involving tretinoin, topical spironolactone in compounded formulations, niacinamide, adapalene, clindamycin with benzoyl peroxide, benzoyl peroxide and azelaic acid. It bars oral drugs, systemic hormonal therapy, isotretinoin and treatment for severe acne.
The system also uses hard stops. Pregnancy or plans to become pregnant, breastfeeding, a weakened immune system, a previous reaction to a listed medication, severe acne and poor-quality images can force the system to reject the automated path or route the patient to a physician. Patients may also request a physician at any point.
If the pilot advances, Stage 2 would let the AI submit prescriptions directly while physicians review every case afterward at least weekly. Stage 3, which requires at least 750 cumulative patients, would reduce routine physician review to a monthly sample of at least 10% of enrollment, while doctors would still review every adverse event and escalation.
Can AI Catch Medication Mistakes?
The safety case for medical AI rests partly on something computers already do well: compare structured information quickly and consistently.
Nolla’s agreement treats a missed drug allergy, contraindication, drug-drug interaction, incorrect dose or frequency, or failure to recommend appropriate diagnostic workup or physician follow-up as a reportable safety event. The company must also report how accurately the system collects medication history and how effectively it screens for drug interactions.
That does not mean an AI system automatically knows every medication a patient takes. The cross-check works only with the data the system can access or the patient supplies. A complete pharmacy record, electronic health record and current laboratory data can make automated screening more powerful. Fragmented records, omitted supplements, prescriptions filled elsewhere or inaccurate patient answers can leave blind spots.
Utah’s separate August AI refill pilot illustrates the broader model. Regulators require that system to verify the patient’s pharmacy record and send the case to a licensed provider when it detects a dangerous interaction, a new side effect or missing laboratory monitoring.
AI can flag a possible problem or a need for further testing, but the quality of the flag depends on the quality and completeness of the underlying data.
The potential benefit matters because medication harm already carries a large toll. The Centers for Disease Control and Prevention found prescribing errors in 7.6% of 1,879 prescriptions reviewed at four Boston primary-care practices. About 3% of all prescriptions contained an error with potential for patient injury, and the researchers estimated that advanced dose and frequency checks could have prevented 95% of those potential adverse events. The study dates to 2005, so it should not serve as a current national error rate, but it shows why automated cross-checking has long attracted interest.
Why Physicians Remain Cautious
Utah’s experiment has already drawn resistance from physicians.
In April, the Utah Medical Licensing Board urged the state to suspend an earlier AI prescription-renewal pilot with Doctronic. The board argued that even routine refills require reassessment for side effects, contraindications, interactions and whether a drug still works as intended.
“There is a reason prescription refills require physician authorization,” the Utah Medical Licensing Board wrote in its April 20 letter to the Utah Department of Commerce.
The state kept the pilot but tightened some controls and continued phased review. Utah has also added independent third-party evaluators to stress-test health AI guardrails, inspect decision logs and assess patient outcomes.
The American Medical Association took a similarly cautious position in June. The organization adopted policy calling for AI to support rather than replace physician judgment.
“AI has enormous potential in healthcare, but it cannot replace physician judgment,” AMA CEO John Whyte, M.D., MPH, said in the June 10 AMA press release.
Nolla points to encouraging early data in its Utah proposal. The company said licensed clinicians agreed with its AI-generated treatment recommendations in about 96% of the last 1,000 treatments and that the disagreements involved minor strength changes rather than different treatment approaches. That figure comes from Nolla’s own operating data, not an independent trial, which makes Utah’s planned outside auditing important.
Could This Happen In Texas?
Texas has already built a legal structure that could support a similar experiment.
House Bill 149, which took effect January 1, created an Artificial Intelligence Regulatory Sandbox Program that expressly includes health care among the sectors eligible for controlled testing.
The law allows an approved participant to test an AI system for up to 36 months without obtaining a license, registration or other regulatory authorization that the state agrees to waive for the test.
The applicant must win approval from the Texas Department of Information Resources and any applicable state regulator, submit a benefit and risk assessment, comply with federal law and file quarterly performance and risk reports. Regulators can remove a participant if the system creates an undue risk to public safety or violates an unwaived state or federal requirement.
Texas also placed guardrails around ordinary clinical use.
Senate Bill 1188, effective September 1, 2025, allows a licensed practitioner to use AI for diagnosis or treatment recommendations based on a patient’s medical record only within the practitioner’s professional scope and requires the practitioner to review AI-created records and disclose AI use to the patient.
Texas law also prohibits an automated decision system from making an adverse insurance utilization-review determination, preserving human control over decisions that can deny or limit care.
Those laws make a limited Utah-style Texas pilot legally plausible, but not imminent.
Texas created the mechanism to waive selected state barriers for supervised testing, yet it also emphasized disclosure, licensed-practitioner review and human responsibility in health care. As of October 6, Texas public materials reviewed by The Dallas Express for this article did not identify a comparable autonomous prescribing pilot.
Are Doctors Taking A Back Seat?
For narrow, repetitive tasks, some doctors may give up part of the front-end work. That does not necessarily mean physicians disappear.
In Utah’s model, physicians define and supervise protocols, handle exceptions, review safety events, answer pharmacists’ questions and remain available to patients. The unresolved issue centers on how much review can safely move from before a decision to after it.
The Dallas Express has previously covered North Texas physicians who described AI as an assistive tool rather than a replacement for medical professionals. Methodist Mansfield, for example, has used AI to flag possible brain abnormalities on CT scans before a neuroradiologist determines whether a patient needs additional treatment.
DX also examined research suggesting that heavy reliance on AI can erode some clinicians’ skills when the technology disappears from the workflow. In another North Texas example, UT Southwestern researchers tested an AI-enhanced electrocardiogram as a screening tool for heart failure risk, with confirmatory testing still playing a central role.
Is This The Future Of Medicine?
Probably in some form, but the near-term future looks less like a robot doctor replacing physicians and more like tightly bounded AI taking over selected low-acuity workflows.
Prescription renewals, screening, image review, documentation and protocol-driven treatment decisions offer obvious targets because software can compare large amounts of structured data quickly and operate around the clock.
The harder questions involve accountability and context.
A doctor can notice when a patient looks ill despite reassuring answers, ask an unexpected follow-up question, reconcile conflicting histories and weigh social or family circumstances that a narrow system may never receive.
AI can also scale a mistake quickly if a flawed rule, biased data set or software update reaches thousands of patients.
Utah’s experiment therefore matters less because AI has already replaced the doctor and more because the state has created a formal path for regulators to decide when a doctor no longer has to approve every routine decision in advance.
Texas has created a sandbox broad enough to ask the same question. Whether either state moves from narrow pilots to mainstream autonomous prescribing will depend on real-world safety data, liability, federal rules, pharmacy participation, physician acceptance and whether patients trust the result.
Provided by Dallas Express
