Daily Briefing for October 6, 2026

Show Notes

Utah lets AI write first acne prescriptions; CommonSpirit credits AI with $200 million; most pathology AI lacks public data; AI intake beats eye residents

The voices on this episode are generated, and so is the script, from the day's post at Physician in the Loop. Every fact in the post links to its source.

Read the post and its sources

Utah approved a pilot Oct. 5 that lets Nolla Health's software issue first-time prescriptions for acne, with physicians reviewing each prescription in advance only for the first 100 patients, MobiHealthNews reported.

The state's earlier sandbox pilots let AI renew existing prescriptions; Nolla's is the first to let AI start a treatment, according to Becker's. STAT reported that Utah also plans to require third-party audits of the claims companies make in its AI sandbox.

Elsewhere, CommonSpirit Health's chief financial officer told investors the system has captured $200 million in financial value from AI. A Helsinki review found no public performance data for 61% of authorized AI products in pathology and hematology morphology, and a randomized trial in China found a language model agent took more complete eye-clinic histories than ophthalmology residents.

In this episode

• Utah lets Nolla Health's AI issue first-time acne prescriptions from a physician-approved list of topical drugs
• CommonSpirit says AI has produced $200 million in financial value while AI spending stays under 0.75% of its IT budget
• No public performance data found for 47 of 77 authorized pathology and hematology AI products, Helsinki review finds
• Language model agent takes fuller eye-clinic histories than ophthalmology residents in 172-patient randomized trial

What is Physician in the Loop?

How technological innovation is changing medicine for patients and the people who practice on the front lines.

Voice 1: This is Physician in the Loop, the daily briefing on AI in medicine. It's Tuesday, October 6th, 2026.

Voice 1: Utah approved a pilot yesterday that lets an app write a patient's very first prescription for acne, something no state sandbox has allowed before.

Voice 2: First-time prescriptions, not refills. What actually has to happen before the AI can send one out?

Voice 1: The company is Nolla Health. It's open only to adults 18 and older with mild to moderate acne who complete an intake questionnaire and a face scan. The app costs about $5 a month, medicines extra, and oral drugs and isotretinoin are excluded.

Voice 2: And a physician still has to sign off on each one?

Voice 1: For the first 100 patients, two physicians review every prescription before it goes out. After that, physicians review the next 500 or more patients weekly for at least eight weeks, then a monthly review covers at least 10% of patients.

Voice 2: The company's chief medical adviser, Dr. Zaid Fadul, put it this way: “The AI can't improvise.”

Voice 1: Clinicians agreed with the AI's recommendations in more than 96% of real-world cases, by the company's own figures. No peer-reviewed results have been published.

Voice 2: Ninety-six percent is a strong number for a count the company ran itself.

Voice 1: It is. Still just their own tally, not an outside one.

Voice 2: How does this compare with Utah's earlier sandbox pilots?

Voice 1: Those, Doctronic and Legion Health, only renew existing prescriptions. This is the first one allowed to start new treatment, and Utah plans to require third-party audits of the claims companies make in the sandbox going forward.

Voice 2: Good, because a self-reported agreement rate is exactly the kind of claim an outside audit should be checking.

Voice 1: Different kind of AI number: CommonSpirit Health's chief financial officer told investors this month that AI has captured $200 million in financial value for the system.

Voice 2: $200 million is a lot. What's it spending to get there?

Voice 1: Michael Browning said direct AI spending stays under 1% of the IT budget, and enterprise token costs run about $25,000 a month.

Voice 2: So a small slice of total spending. Who's deciding what actually gets deployed?

Voice 1: A 36-member governance committee has reviewed more than 122 enterprise AI uses and rejected 20 over bias, safety risk or vendor overlap. The clinical ethics team holds a veto over any deployment.

Voice 2: Did Browning say what's actually driving that $200 million figure?

Voice 1: No, the system didn't break it down by tool. He did say health care remains, in his words, “roughly 80% people, 15% process and only 5% technology,” and that AI doesn't change that ratio.

Voice 2: A tidy breakdown from a CFO citing a nine-figure savings number in the same call.

Voice 1: A different kind of accounting problem showed up in a review out of Helsinki. Researchers at Helsinki University Hospital and the University of Helsinki looked at 77 authorized AI diagnostic products in pathology and hematology, and found public performance data for only 30 of them, that's 39%.

Voice 2: So 47 out of 77 are approved with nothing public to check. Is that spread evenly across categories?

Voice 1: No. Evidence was available for all eight products authorized in both the European Union and the United States, but for only 34% of pathology products, against 78% of hematology morphology products.

Voice 2: So hematology morphology is far better documented than pathology.

Voice 1: It is. And even the published evaluations used 15 different performance metrics, which the authors say limits comparing one product against another.

Voice 2: Fifteen different yardsticks makes it hard to say any one product is actually the best.

Voice 1: Which is close to the point they're making. Did they say why so much of this stays private?

Voice 2: They note performance evaluation results simply don't have to be published to get authorized. Their phrase for it was “a gap between evidence availability and regulatory authorization.”

Voice 1: A trial out of Guangzhou looked at who takes a better history before an eye appointment, a language model or a resident.

Voice 2: And the AI won?

Voice 1: In a 172-patient randomized trial at Zhongshan Ophthalmic Center, the language model agent scored nearly 18 points higher on a history-quality scale after adjustment.

Voice 2: Patients liked it too, from what the post says.

Voice 1: They rated it higher for empathy, yes. It also took longer, a median of about 11 minutes against roughly 3 for the residents.

Voice 2: Did the extra time translate into better diagnoses?

Voice 1: From the history alone, yes: an F1 score of 0.85 versus 0.68 for the residents. Once a physical examination was added, it helped the residents' accuracy more than it helped the agent's.

Voice 2: So the edge shows up in the conversation, not the exam room. The trial ran at one hospital in 2025, and the authors say it may not carry over to other specialties or settings.

Voice 1: Sources and corrections are at physicianintheloop.org. This briefing was read by AI voices.