The AI doctor's first credential is a billing code

The figure is 60 to 80 percent. That is what federal officials have discussed paying company-run "AI physicians," as a share of the human fee for the same service, the New York Times reported on Monday. Nothing has been proposed, and John Whyte, the AMA's chief executive, said in the same story that models built to provide care on their own are not ready for widespread use. Both caveats are fair, and the story is still the most consequential thing written about this profession all year, because in this country nobody gets replaced until someone gets paid. The rate, though, is a rumor. The contract Medicare expanded on Tuesday is a signed document, and it requires no AI to be licensed, cleared, or even especially accurate, only a company to be paid for your patient's blood pressure while you are paid per visit for the same patient.

That contract is the ACCESS model, which pays technology companies for outcomes in hypertension, diabetes, musculoskeletal pain, depression, and anxiety, and which will add heart failure, COPD, substance use, and tobacco next spring. CMS says three out of four people with Medicare now qualify for at least one track, and that payers covering 165 million people with Medicare Advantage, Medicaid, and commercial plans have pledged to copy the structure; if they follow through, it lands in your contracts in January. On Thursday Counsel Health, which sells an AI front door with a $29 physician follow-up, said it will join ACCESS in early 2027 with Oura as its wearable partner, treating hypertension, obesity, hyperlipidemia, and prediabetes at no cost to the beneficiary and promoting the offer inside an app with about five million paying members. Counsel reports 36,000 Oura members onboarded in nine weeks and a care rating of 4.6 out of 5. Those are the vendor's numbers, and I file them as such.

The case for ACCESS is serious and deserves its full strength. Medicare has paid for visits and got visits; paying for a controlled blood pressure instead is on its face the better bargain, eighteen clinical and patient societies back the model, and a beneficiary treated at no cost by a company paid for her outcomes is not obviously worse off. Grant all of it, and ask what evidence the companies have. A paper in npj Digital Medicine looked at 229 randomized trials of digital health tools and found the developer involved in 73 percent of them, with somewhat higher odds of a positive result. The effect is modest and the domains are behavioral, but that is exactly the evidence the ACCESS participants are selling on.

The rest of the answer came from the same agency, which spent the week explaining why its AI gatekeeper is failing. About a thousand pages of records that the Electronic Frontier Foundation won in a public-records suit show one WISeR vendor telling CMS before launch that it would auto-affirm requests because it was not ready, another, Virtix, placed on a corrective action plan after denying more requests than it approved, vendors paid per denial with quality penalties of only 5 to 10 percent, and a request that sat for 83 days. A payment design shapes a vendor's conduct from the first day, long before anyone learns whether the tool works: pay per denial and you get denials. ACCESS pays for outcomes in your patients, so the first question to ask is whether you will see the data.

On Wednesday Chris Klomp, the nominee to run HHS day to day, sat before the HELP Committee while Senator Patty Murray read those numbers to him. He first said the contractors were not paid by denial volume, then allowed that there are penalties for inappropriate denials, and then said the program "must be done appropriately, or it should not expand." Ms. Murray said she would fight the planned move into oncology. A condition is weaker than a stop, but the person who will oversee the program stated it on the record, and it fits in an appeal letter.

The terms can be written the other way, as Connecticut showed the same day. Sean Scanlon, the state comptroller, set five rules for the 270,000 people on the state employee and municipal plans: no adverse determination made solely by an AI, no AI-only downcoding or payment reduction, disclosure when AI is materially involved in a benefit decision, no training of other models on members' data, and validation audits. He wants the legislature to extend them to every state-regulated plan in 2027, but they are plan policy rather than statute, so any large employer plan, including a health system's own, could copy them next month, and should: the two dozen revenue cycle leaders Becker's polled this week said what they fear is not autonomous coding but payers reviewing 100 percent of claims with automation and quietly reducing payment rather than denying it. Sacramento, meanwhile, has said nothing about AB 1979, the first state bill written against the staffing model itself: no AI performing licensed work, no AI directing unlicensed staff to do it. The governor has until September 30.

The tool those contracts will buy is also taking shape. The best paper I read this week, in Nature Medicine, comes from Jakob Nikolas Kather's group, whose diagnostic agent, built on open-weight models running inside the hospital firewall, took histories, ordered tests, and committed to diagnoses on 551 emergency cases across seven conditions. Accuracy was about 90 percent, close to a cloud model. The useful result came from running it five times and asking whether it agreed with itself: at a strict consistency threshold it kept 49.4 percent of cases at 98.9 percent accuracy and routed the other half to a person. One week after ARPA-H put $62.7 million into heart-failure agents built on exception-based oversight, that is the mechanism written down. Under it the cases that reach you are the hard ones by construction, so your time per case goes up, and any productivity target that assumes otherwise is wrong on arrival.

The money kept moving in one direction. Sword Health is buying Headspace for a reported $300 million in cash, about a tenth of what Headspace was worth in 2021, to fold it into an AI that decides when to bring in a clinician. Tenet's Conifer will cut 1,037 billing jobs by November 2, with the WARN notice citing technology initiatives. Waystar hired bankers to explore a sale after the market decided AI would eat billing software. Becker's counts ten health systems that cut IT and coding roles this year, and not one named AI. There is still no verified case of a US health system cutting physician positions because of AI. What there is, per Doximity, is 20 percent of physicians already facing higher productivity targets because of it. Replacement arrives as a bigger census rather than a pink slip, which is harder to fight because nobody announces it.

So argue about the contract. If you practice in Arizona, New Jersey, Ohio, Oklahoma, Texas, or Washington, put Mr. Klomp's sentence and the EFF numbers in your next WISeR appeal. If you sit on a contracting or medical society committee, ask two questions before the January contracts are signed: which ACCESS-aligned companies will be paid for outcomes in your patients, and whether you will see their data. If your health system insures its own employees, hand the benefits office Mr. Scanlon's five rules for adoption next month. If you bill remote monitoring through a vendor, price the employed-staff version before the final fee schedule lands around November 1. If your system starts saying exception-based review, ask what fraction of cases the agent keeps and how accurate it is on the ones it sends you, and get the answer into the productivity conversation before the tool arrives. And when anyone hands you a study, ask who ran it; in 229 trials, 73 percent of the time it was the company. The billing code is being written. The examination is still yours to give.