Doctors sit an exam. The agent signs a contract.

Sam Toole and Hannah McQuaid, two partners at the venture firm Primary, published an essay in July called "The next medical provider." Their claim is that agentic AI is the next one through the door after nurse practitioners and physician assistants, "the next iteration of a new 'medical provider' entering the system and creating leverage." Their picture is a primary-care physician whose panel "caps out around 1,800 patients, if not much lower," stretched to about 2,500 with ambient scribes and triaged inboxes, and then to "12,000+ lives," with the physician "cosigning 200 AI-drafted encounters a day, seeing only 3-4 escalation visits." The 1,800 holds up against survey data on family physicians. The rest is the authors' own projection, and they show none of the working.

Every provider type before the agent came in the same way. A physician gets a state license after passing the USMLE, the exam whose results go to state boards "for use in granting the initial license to practice medicine." A physician assistant passes the PANCE, and the certification it earns is a requirement for licensure in every state. The new provider sits an exam about the work, and then a state lets it do the work.

The agent has no exam. Reading Primary's piece, I took that for an oversight, the kind somebody gets around to fixing once the money is large enough. I had it backwards.

The exam was refused

The Federation of State Medical Boards co-owns the USMLE with the National Board of Medical Examiners. In August its president and the chair of its board wrote in STAT that "AI is not ready to be independently licensed like a physician." Boards, they wrote, "regulate the physicians who use them," and "professional responsibility remains with the licensee." Idaho and Iowa each had a bill this year for a separate board to oversee AI, and both failed. Delaware went further in April and put it in statute (as Holland & Knight summarises HB 191): a nonhuman entity, AI agents included, cannot be licensed as a physician, nurse or physician assistant, or use the titles.

So the people who own the exam looked at the candidate and declined to seat it. That is a decision, and a defensible one. A license certifies a person who will hold it for forty years, can lose it for misuse, and can be put under oath. A model its vendor can swap out next quarter is none of those things.

When I wrote about AI workers being slowed down by law early last year, the open question was whether the medical board would license the AI. The boards, and one legislature, have answered, and the answer is no. Most of that piece has held. Work outside the licensed scope did ship first: WellSpan's voice agent now takes more than 160,000 patient calls a month, mostly about appointments. Mental health moved faster than I expected, in the restrictive direction, with Illinois, Maine and Tennessee all limiting AI in therapy since August 2025. What I missed was a third route, one that goes around the board altogether.

The exam became a contract

The route runs through Utah. The state's Office of Artificial Intelligence Policy, which sits inside the Department of Commerce, can grant a company "regulatory mitigation" for a year at a time. In January it gave one to Doctronic, to renew prescriptions with an AI. The agreement does the job an exam would do, per product instead of per practitioner. For the first 250 patients, a licensed physician reviews every renewal before it goes to the pharmacy. For the next 1,000, every renewal is reviewed after the fact. After that, physicians review a sample of 5 to 10% each month, and the results go to the state medical board. The phases come from the safety protocol the company submitted with its application, so in Utah the grader built the test as well. In October the same office gave a second company room to write first prescriptions.

The agreement is also plain about what it is not. It "does not constitute an endorsement or approval" from the state. And in Section 8, Doctronic agrees to hold the office and the state's licensing division "harmless from any claims, liabilities, damages, losses, or expenses" arising from its work under the waiver. The copy the state posted is headed DRAFT, which is the one caveat on everything in this paragraph.

Put the two routes side by side and there are now two answers to the question a license used to settle, which is who carries the harm when the care goes wrong. Under the boards, in Texas and in Illinois, the licensee carries it. Texas allows AI-assisted diagnosis only inside a practitioner's scope and makes the practitioner review the records the AI produced. Illinois bars AI from making therapeutic decisions without review by a licensed professional. Under Utah's waiver the contract points the other way: the vendor holds the state harmless, and no clause in the agreement names a physician as the one who absorbs the harm.

A license names who to blame. A contract names who pays.

Two hundred signatures a day

Primary's scenario lands squarely in the first regime. A physician cosigning 200 AI-drafted encounters is the licensee on all 200, and the responsibility the FSMB described stays with her for each one. In an eight-hour day with nothing else in it, 200 signatures is 2.4 minutes a chart, before the three or four escalation visits the scenario books. I could not find a state that has written down what supervising an agent means. Texas comes closest, with a practitioner who reviews the records. At 2.4 minutes, that review is the thing a plaintiff's lawyer will test, and the doctrine that the human in the loop carries the harm will hand it to her.

Primary's authors see the liability question coming. They list malpractice among the barriers and note that today's copilots "explicitly defer liability to the clinician." Their scenario takes that deferral, multiplies it by two hundred a day, and gives all of it to one person.

Nor is there an informal exam waiting to stand in. Mercor mapped 29 public AI benchmarks onto the tasks of 923 occupations and found that only 12 of 89 healthcare practitioner occupations had any task covered by at least five prompts. Mercor calls the results provisional and probably low, and the map measures how much of the work anyone has tested, not how well a model does it. For most clinical work, then, no public test would show it either way.

A routing machine with a doctor at the end

The supply framing (one physician, almost seven times the patients) hides what the agent in that scenario does all day. Of everyone who comes in, it decides which three or four see a human. That is a routing decision, made across twelve thousand lives. I argued in 2024 that Ontario's primary-care shortage was a routing problem more than a supply problem, and at Medimap the product was routing, getting a patient to a walk-in clinic that could see them that day. Primary's scenario makes the same case from the other side. Most of what the agent is worth is in the sorting.

And the sorting is the step nobody examines. Utah's phases at least count it: the agreement benchmarks how often the AI approves a renewal and how often a patient or a pharmacist asks for a physician instead. In the licensee regime the triage arrives pre-sorted. The physician signs 200 drafts and sees three or four patients, and whether the right three or four got through is a question her signature answers without anyone asking it.

The exam was refused on purpose, by the people best placed to write it, for reasons that hold up. What has grown in its place is a test each company proposes for itself, one state and one product at a time, filed with a commerce office, plus a clause saying who pays when it fails. Every provider before the agent was examined as a person. This one is being admitted as a vendor.

—TJ