This analysis synthesizes 9 sources published the week ending Nov 10, 2026. Editorial analysis by the PhysEmp Editorial Team.
HCA Healthcare expects about 10,000 of its physicians to be using ambient documentation AI by the end of this year, and Doximity says more than 190 health systems have now bought its clinical AI suite, up from 165 a quarter earlier. The tests used to judge these tools describe a different job. A preprint covering 127,833 clinician conversations with one academic system’s AI assistant found that the median published medical AI benchmark overlapped with only 31 percent of what clinicians actually asked it to do.
For AI in Physician Employment & Clinical Practice, that mismatch matters because the validation the benchmarks skip does not disappear. It moves to the physician using the tool, during the visit, on time nobody has budgeted.
Scale arrived before the yardstick
This week’s rollout numbers came from the organizations doing the rolling out. HCA chief clinical officer Michael Schlosser, MD, told Becker’s the tool will reach more than 100 of the system’s roughly 189 hospitals for hospitalists and emergency physicians by year-end, plus about 3,000 outpatient physicians and APPs. He put the savings at about five minutes per note, or one to two hours per shift, and said outpatient clinics have added patients to schedules. Doximity CEO Jeff Tangney told investors the company’s AI suite now sits in 10 of the nation’s top 20 hospitals.
Physician uptake is moving more slowly than vendor footprints. A Doximity survey of 3,151 physicians, cited by PYMNTS, found ambient scribe use rose from 20 percent to 29 percent between two survey waves. Asked about their institution’s AI decision-making, 47 percent called it still evolving. Only 8 percent called it clear.
Even a Doximity researcher concedes the sequence is off. In a Doximity-partnered Modern Healthcare conversation, chief AI researcher Louis-Antoine Mullie, MD, a former intensivist, said that with some AI tools “deployment has come first.” His proposed minimum standard is plain: a physician using AI should provide safer care than a physician without it.
What the benchmarks test, and what clinicians ask
The preprint, summarized by Telehealth.org and not yet peer-reviewed, examined the first message in conversations from 6,342 physicians, APPs and nurses across 35 specialties between August 2025 and April 2026. Documentation and administrative requests made up 36.2 percent. Knowledge retrieval was 28.9 percent. Diagnosis was 3.7 percent. The median of the 58 public benchmarks compared contained no documentation requests at all. NOHARM, the safety benchmark Doximity cites when it reports ranking first, classified 54.5 percent of its items as high or critical risk, against 13.2 percent of real clinician queries. The study says this does not invalidate NOHARM.
The more practical finding concerns incomplete questions. About 36.1 percent of clinician requests could not be answered as written, either without substantial assumptions or at all. Among high-risk requests, 84.5 percent called for clarification or another response rather than a direct answer. Benchmarks hand a model a complete case. Clinicians at work rarely do.
How a question is packaged can swing results too. Pediatrician Vinson James, MD, writing in KevinMD, described testing six models on 77 pediatric board-style questions. One model scored 38 percent when the questions arrived in a single bulk prompt and 83 percent when they came in batches of 10. He wants institutions to benchmark tools “under realistic clinical load, with standardized prompts, before go-live.” None of this week’s deployment announcements described doing that. Looking back over our own recent coverage, we have mostly stayed out of the public argument over whether AI already outperforms physicians, and a week spent questioning the measuring sticks suggests that debate deserves more of our attention.
A strong showing on a consult-style safety test says little about a tool that most clinicians use to draft notes and look things up. Systems that buy on those rankings are handing the local validation step to their own medical staff.
The check lands on whoever signs
Rhode Island has now written that check into law. Its new notification act, per PYMNTS, requires providers who use AI to document visits to tell the patient and to review the AI-written note. Louisiana’s Act 649 requires verbal disclosure before AI transcription begins, with licensing-board discipline possible for skipping it. Neither statute says how long review should take or who pays for the time.
Evidence on how well physicians catch bad output is thin and unsettling. Gastroenterologist Brian Hudes, MD, describes a medRxiv preprint in which 225 physicians worked written neurocritical-care cases with AI advice attached. When the advice was deliberately steered toward harm, physicians made a harmful decision 49.4 percent of the time, versus 2.0 percent with aligned advice, and only 2 of the 225 detected the manipulation. Hudes, who discloses that he founded an AI company, warns against carrying that figure into an exam room: the cases were written, one model was tested, and language models judged the outcomes. His structural point stands regardless. Malpractice carriers pay for failures, he writes, not “the review minutes that might prevent them,” and the AMA’s new governance framework is silent on who funds oversight.
Bedside teams already run their own filter. In a Frontiers in Digital Health field study of 14 cardiothoracic ICU team members, one participant estimated that about 85 percent of alarms were false. An EHR sepsis tool went largely unused, partly because its notifications arrived hours late. Asked what confidence level would make a prediction useful, participants named anything from 75 to 100 percent and never converged.
Where uncounted review time goes
When sanctioned tools are narrow, clinicians improvise. Emergency physician Harvey Castro, MD, points to a reported incident at Johns Hopkins Bayview in which an employee uploaded a spreadsheet covering nearly 3,200 patients to an unapproved public AI tool. His explanation is friction: the approved tool was bought for one job, usually notes. Tangney expects health systems to consolidate around approved applications as governance tightens. Consolidation limits which tools physicians may use. It does not tell anyone how those tools perform on a given service line.
HCA’s own account of the time saved already shows where some of it went: more patients on outpatient schedules. If saved minutes become volume before anyone counts review time, the review happens inside a day that no longer has room for it.
For physicians weighing offers, the questions worth asking are now specific. Which AI tools are required, what local testing preceded go-live, whether productivity targets changed after deployment, and whether review of AI-drafted notes or advice counts anywhere in the compensation formula. Recruiters at systems with clear answers hold an advantage that is easy to quantify, since only 8 percent of surveyed physicians could describe their own institution’s AI decision-making as clear.
Two dates frame the next year. HCA plans to finish its rollout by the end of 2027, and Rhode Island’s review rule is already in force. If more states copy the review mandate before employers attach time or credit to it, the first contract disputes over clinical AI will be fought over hours rather than liability.
Sources
HCA Healthcare to have 10,000 physicians using ambient AI by year-end – Becker’s Hospital Review
Doximity expands clinical AI footprint to more than 190 health systems – Fierce Healthcare
As AI Listens in on Doctor Visits, States Are Writing the Rules – PYMNTS
What Should Clinical AI Have to Prove Before We Trust It? – Modern Healthcare
Study Finds Gap Between Medical AI Benchmarks and Clinicians’ Real-World Use – Telehealth.org
Clinical AI scored 38 and 83 percent on the same questions – KevinMD
Who pays physicians to catch AI medical errors? – KevinMD
Clinicians in the Cardiothoracic ICU Say AI Must Earn Its Place Before They Will Trust It – Bioengineer.org
Shadow AI: why clinicians paste patient data into chatbots – KevinMD