Since October 5, Utah residents 18 and over can download an app called Nolla Derm: verify your identity, sign an informed consent, answer a short intake questionnaire, photograph your face from five angles, and 10 to 15 minutes later you have an acne treatment plan. If you qualify, a prescription goes out for home delivery or pharmacy pickup. Pilot price: $4.99 a month. No doctor visit anywhere in the patient’s side of the flow (Nolla’s announcement).
The tech-press version: “Utah will become first state to let AI examine patients and prescribe medication without direct human oversight” (TechSpot). I went and read the state’s own documents. Two facts change that story.
First, under the agreement’s phase one, every AI-generated prescription is independently reviewed and approved by two Utah-licensed physicians before it reaches the patient (Utah Department of Commerce pilot page). Second, Utah’s earlier AI pilot, Doctronic, which handles only prescription renewals, signed in October 2025 and has run for nearly a year without written approval to enter phase two. A licensed practitioner still authorizes every renewal, and the state amended the agreement mid-course to tighten the threshold and remove two medications from the list (same pilot page).
“Unsupervised” is the endpoint of this design, not its current state. And judging by Doctronic’s pace, the endpoint is not guaranteed.
Who is actually prescribing
Nolla is not a company improvised for a stunt, but it is not physician-led either: the licensed doctors sit in its advisory layer, not its founding team. It launched in Norway with a skin cancer screening app in November 2024, added an acne app in May 2025, and reports serving over 55,000 patients across both within six months, roughly 1% of Norway’s population (Beauty Independent). In 2025 it expanded to 40-plus US states under an AI-assisted, clinician-reviewed model, with a $4.5 million seed round led by General Catalyst (General Catalyst).
Distribution is direct-to-consumer: users download from the App Store, no physician referral, and I found no partnership announcements with hospitals or telehealth platforms like Teladoc. The scale figures are the company’s own: 175,000-plus downloads across 44 states (Nolla’s announcement). Downloads, not paying patients. The underlying market is large either way: acne is the most common skin condition in the US, affecting up to 50 million Americans a year (AAD statistics). The pilot’s most direct acquisition lever is price. Nolla’s physician-reviewed service costs $59 a month in other states (Beauty Independent); the Utah pilot is $4.99, less than a tenth. Whether there are ad-spend or channel plans behind it, the company has not said, and I could not find out.
The founder closest to medicine is Luis Ruben Soenksen: trained in biomedical engineering, with a bioengineering master’s from Johns Hopkins and a mechanical engineering PhD from MIT, formerly MIT’s first venture builder for AI in healthcare (MIT Proto Ventures bio), with what investor General Catalyst describes as a decade of research on dermatological AI modeling. He is an engineer and researcher; Nolla’s team page lists him as founding scientific advisor, not a physician. CEO Luis Wenus comes from growth operations, previously Worldcoin’s first employee leading go-to-market; the third co-founder is Sean Geiger. The physicians on the team page are all advisors: chief medical advisor Zaid Fadul, a family medicine physician, plus Harvard dermatologist Omar Badri and epidemiologist Michael Mina (Nolla team page). The agreement allows the reviewing physicians to be either employed or contracted; which arrangement Nolla actually uses is not stated in the official materials.
Why the state can authorize this
The legal basis is SB 149, Utah’s 2024 AI Policy Act. It created the Office of Artificial Intelligence Policy (OAIP) and an “AI Learning Lab”: a company can sign a regulatory mitigation agreement with the state, under which Utah suspends enforcement of specific regulations within a defined scope and term, and in exchange the company accepts the agreement’s full set of oversight conditions. A regulatory sandbox: fence off a small area, let the new thing run, collect data as it goes, instead of legislating for every case up front.
Why a state rather than the FDA? Because jurisdiction already splits this way: the FDA regulates drugs and devices themselves and does not regulate the practice of medicine (FDA). Who may examine patients and prescribe is a matter of state licensing law. Utah waived exactly the two levers it owns, enforcement of unlicensed practice and telehealth requirements, and only where the AI issues a topical acne prescription under the agreed protocol.
The AI did not get a blank check. The scope is narrow: mild-to-moderate acne only, a fixed list of topical creams and gels, isotretinoin and all oral medications explicitly prohibited. Severe acne, pregnancy or plans to conceive, breastfeeding, a compromised immune system, or a history of reactions to the listed drugs each knock a patient out of the flow.
What the state picked was not “dermatology” but one specific scenario, topical acne, which satisfies conditions that almost no other prescribing scenario does at once. Acne is diagnosed primarily by looking at lesions (AAD treatment page), which suits photo-based assessment. Topical dosing produces far lower systemic exposure than oral drugs, and the common adverse reactions, dryness and local irritation, are ones the patient can see for themselves. Narrow is not zero risk: the same AAD page notes that folliculitis and perioral dermatitis can be mistaken for acne and are treated differently, and the FDA label for topical clindamycin warns about percutaneous absorption and rare but potentially fatal pseudomembranous colitis (FDA label). But the highest-risk part of the specialty is carved out wholesale: isotretinoin’s teratogenicity, with FDA-mandated pregnancy testing and prevention under iPLEDGE, is the first item on the prohibited list. Try the same design on initial prescriptions in another specialty, where drugs act systemically and diagnosis depends on physical exams and labs, and almost none of these conditions hold.
Oversight runs in three phases. Per the signed agreement (agreement text): for the first 100 patients, two physicians review every prescription before it is sent. From there to 500 cumulative patients, the AI sends prescriptions directly and physicians retrospectively review all cases at least weekly. Above 750, physicians audit at least 10% of cases monthly plus every adverse event. Each advancement requires written OAIP approval; Nolla’s stated bar includes a 95% AI-physician agreement rate and zero serious adverse events. On top of that: adverse events reported to the state within 24 hours, monthly concordance data on how often the AI and reviewing physicians agreed, mandatory professional liability insurance, no liability caps in the terms of service, no selling patient data. A patient’s right to sue is untouched.
One thing I could not verify at clause level: once prescriptions go out AI-first and physicians move to retrospective review and sampling, how the malpractice standard of care applies. The official summary does not expand on it.
This is not a one-off waiver
Utah announced four more agreements the same day: August AI for chronic-disease refills (first 250 refills fully reviewed, at most three consecutive AI renewals before a patient must see a human, and any sign of suicidal thinking or a new side effect stops the refill and routes it to a licensed provider); Expect Fitness for postpartum pelvic-floor physical therapy; and umbrella agreements with University of Utah Health and Intermountain Health that give their future AI pilots an expedited review path. The state also retained six third-party evaluators for independent technical and clinical review, including the Coalition for Health AI and Stanford’s Clinical Excellence Research Center (Commerce press release).
Nolla’s own claim is “nation’s first.” Some outlets inflated that to “world’s first” (Interesting Engineering), and that needs splitting in two. “AI examines, human signs” has precedent: Shanghai-based Synyi AI has been running a clinic in Saudi Arabia’s Al-Ahsa region since April 2025, where its AI doctor “Dr. Hua” conducts consultations and drafts treatment plans across roughly 30 respiratory conditions, with every prescription requiring a human physician’s sign-off (Bloomberg). Inside China, hospital AI is widely deployed, nearly 450 medical AI products in Shenzhen alone according to a CCTV report (republished by the Cyberspace Administration), but regulators closed the prescribing step outright: the 2022 internet-diagnosis rules require the treating physician to issue prescriptions personally and expressly prohibit AI-generated prescriptions (National Health Commission, official repost).
Utah’s increment is the second thing: a regulator put in writing a path by which physicians step back from prospective review of every script to retrospective audits and sampling, and it covers initial prescriptions. On that specific point, initial prescriptions, my public search turned up no earlier equivalent authorization. Utah’s own Doctronic agreement wrote down a similar path a year before Nolla’s, but for renewals only. Note also that right now, with every Nolla prescription still passing two physicians, the setup is in practice no less supervised than the Saudi clinic’s; if anything more so, with two reviewing physicians to the Saudi clinic’s one. The real divergence begins only if phase two is approved.
In July I wrote about how HIPAA attaches to institutions, not to data: standalone consumer AI apps usually sit outside it, and records handed to them leave the protected zone. Medical licensing has the same structural problem. It attaches to people, not systems. An AI holds no license to revoke, so the regulator’s existing tools get no grip on it; that is precisely why Utah had to waive unlicensed-practice enforcement (pilot page). The state’s substitute is a contract that can be pulled: miss the concordance bar and the next phase is not approved; something goes wrong and the agreement terminates, enforcement resumes. The constraint moved from credential gatekeeping to continuous auditing of data.
What is actually being tested
Not AI’s ability to read acne. Topical acne was picked to push clinical risk to the floor so that something else could face the test: whether contract-plus-data-thresholds can function as a replacement for a license.
The most convincing evidence so far is Doctronic sitting in phase one. The gate actually holds people back; advancement is not a formality. That says more about the pilot’s seriousness than any adjective in a press release.
Two weak points are worth watching. One is capacity. OAIP now holds multiple agreements, active pilots plus umbrella frameworks, each generating monthly reports, concordance data, and incident filings to check. I could not find staffing figures for the office in its public materials. If agreements multiply faster than staff, “continuous auditing” can quietly degrade into “collecting data for the archive.” The other is replication drift. These agreements are negotiated case by case, not statutes. A state that wants the headline can copy “AI prescribing approved” without copying the two-physician review or the written-approval gates. That is the point at which “unsupervised AI prescribing” stops being a headline error and becomes true.
One concrete date to mark: Nolla’s agreement expires October 5, 2027. Whether it has reached phase three by then, and whether the AI-physician concordance data gets published, will matter far more than the news that a state went first.
References
- Utah Enhances Pro-Human AI Initiative with New Healthcare Pillar… (Utah Department of Commerce, 2026-10-05) — healthcare pillar, five new agreements, third-party evaluators, 24-hour reporting and insurance requirements
- AI Authorized Pilots (Utah Department of Commerce) — scope, exclusions, phase structure, and waived enforcement for the Nolla/Doctronic/August AI agreements; Doctronic’s phase-one status and the tightening amendment
- Regulatory Mitigation Agreement with Nolla Health (Utah Department of Commerce, PDF) — phase thresholds (100/500/750), review cadence, advancement criteria, insurance and liability terms
- SB 149 Artificial Intelligence Amendments, enrolled text (Utah Legislature) — statutory basis for the AI Policy Act, OAIP, Learning Lab, and regulatory mitigation agreements
- Utah and Doctronic Announce Groundbreaking Partnership (Utah Department of Commerce, 2026-01-06) — official positioning of the Doctronic pilot
- Nolla Health Launches the Nation’s First AI-Powered Prescriptions (Nolla Health) — patient flow, pilot pricing, download and state-coverage figures, phase descriptions and advancement bar
- Seeding the Future with Nolla Health (General Catalyst) — seed round, founder backgrounds, US expansion, Soenksen’s decade of dermatological AI research
- With $4.5M Raise, AI Startup Nolla Health Offers Acne Care (Beauty Independent) — Norway app timeline (November 2024 and May 2025), 55,000 patients in six months, $59 monthly US price, the three co-founders
- Our Team (Nolla Health) — medical and scientific advisor roster
- AI + Healthcare channel, Luis Soenksen bio (MIT Proto Ventures) — Soenksen’s degrees and MIT venture builder role
- Skin conditions by the numbers (American Academy of Dermatology) — acne as the most common US skin condition, up to 50 million Americans annually
- Acne: Diagnosis and treatment (American Academy of Dermatology) — visual diagnosis of acne; folliculitis and perioral dermatitis as look-alikes
- Cleocin T topical clindamycin label (FDA) — percutaneous absorption and pseudomembranous colitis warnings
- iPLEDGE REMS (FDA) — isotretinoin teratogenicity and pregnancy-prevention requirements
- FDA’s Role in Regulating Medical Devices (FDA) — FDA does not regulate the practice of medicine
- Chinese Startup Trials First AI Doctor Clinic in Saudi Arabia (Bloomberg, 2025-05-15) — Synyi AI’s Saudi clinic flow, disease scope, human sign-off
- Internet Diagnosis and Treatment Supervision Rules (Trial), release page (National Health Commission of China) — China’s prohibition on AI-generated prescriptions (original release)
- Same rules, official repost (Guangzhou municipal government portal) — clause text cross-checked against this repost
- Jiaodian Fangtan report on AI in healthcare (CCTV, reposted by the Cyberspace Administration of China) — deployment scale of medical AI in China (nearly 450 products in Shenzhen)
- Utah launches world’s first autonomous prescription system without human doctors (Interesting Engineering) — example of the “world’s first” framing
- Nolla Health Launches AI-Issued Initial Acne Prescriptions in Utah (Unite.AI) — three-phase thresholds (100/500/750, 10%), company’s “first” claim
- TechSpot report — the headline this article checks against the agreement text