Priced before proven

Federal officials have discussed paying company-run "AI physicians" as much as 60 to 80 percent of what human doctors earn for the same service, the New York Times reported on September 14. The officials are still exploring the idea, and John Whyte, the chief executive of the American Medical Association, said that models designed to provide care on their own are not ready for widespread use.

Both caveats are fair, and the report still matters more than any benchmark published this year, because in American medicine nobody is replaced until somebody is paid. The rate is only talk, but the contracts that decide how machines are paid to look after patients are already being written, before anyone knows how well the machines work. Those terms will shape what the technology does to medicine far more than its accuracy will, and doctors should spend their energy fighting over them.

The most important contract is ACCESS, a Medicare program that pays companies using technology to manage chronic disease for measurable improvements, such as a patient's blood pressure brought under control, rather than for each service. On September 15 Medicare said it would add heart failure, chronic lung disease, substance use disorders and tobacco cessation next spring, that three in four people on Medicare qualify for at least one of its tracks, and that insurers covering 165 million Americans had pledged to adopt a similar way of paying for results. Two days later Counsel Health, which pairs AI with its own physicians, said it would join in early 2027, treating high blood pressure, obesity, high cholesterol and prediabetes at no out-of-pocket cost, with Oura, the maker of a smart ring, promoting the offer to its own paying members.

Medicare's other experiment shows what such terms do. WISeR, a pilot in six states, pays private companies that use AI to review requests for prior authorization, the advance approval a patient needs before some treatments. Records obtained by the Electronic Frontier Foundation show that the companies are paid for the requests they deny, that poor quality scores trim that pay only slightly, and that one company had to file a corrective plan after denying more requests than it approved. Pay for denials and you get denials. At a Senate hearing on September 16 Chris Klomp, the nominee for deputy secretary of health and human services, was asked whether the companies make more money when they deny care. "My understanding is no," he said, though he later agreed that the program "must be done appropriately, or it should not expand."

The evidence the vendors sell on is thinner than their pitch. A review in npj Digital Medicine of 229 randomized trials of digital health tools found the developer involved in nearly three-quarters of them, and those trials were somewhat more likely to report a statistically significant result. Tools already running in hospitals can disappoint when outsiders test them. A study in JAMA Network Open ran Epic's end-of-life score, which estimates who will die within a year, on hospital patients at two large health systems, and found that it ranked them reasonably well but overstated the risk of death among sicker patients and grew less reliable with age.

The machines those contracts will buy are taking shape. A diagnostic agent described in Nature Medicine, running on open models inside a hospital's own computers, worked through each emergency case repeatedly and kept only those on which its answers agreed; it was right about 99 times in 100 on the half it kept, and passed the rest to a doctor. Under such designs the cases that reach a physician are the hard ones by construction. Yet a fifth of doctors in a Doximity survey say they already face higher expectations for productivity because of AI, and a list of this year's job cuts at health systems, in IT, coding and other back-office work, does not mention AI at all. Replacement is arriving as a harder day's work rather than a layoff notice, and nobody announces it.

The strongest objection is that paying for results beats paying for visits. Medicare has bought visits for decades and got visits; a company paid for lower blood pressure has every reason to lower it, eighteen clinical and patient societies back ACCESS, and a patient treated at no charge is not obviously worse off. All of that may be true, and all of it depends on the terms: who measures the result, who sees the data, and what happens to the patients whose numbers do not move. WISeR shows how quickly a payment formula becomes a vendor's behavior.

The terms can be written the other way. On September 16 Sean Scanlon, Connecticut's comptroller, set rules for the state's health plans for public employees, which cover more than 270,000 people: no AI may deny care or cut a doctor's payment on its own, insurers and providers must disclose when AI materially assists with or recommends a member's benefits or care, and members' data may not be used to train other AI models. He will ask the legislature next year to extend the rules to every plan the state regulates.

Medicare should publish how often each WISeR company denies care and how often appeals overturn those denials, and should demand trials run by someone other than the vendor before it pays for care by machine. Physicians in the six WISeR states should quote Mr. Klomp's words back to Medicare in their appeals. Other states, and employers and hospitals that insure their own staff, should adopt Mr. Scanlon's rules. Anyone setting productivity targets for doctors who supervise AI should count the hard cases the machines send back, and every doctor handed a vendor's study should ask who ran it.

The machines have been priced before they have been proven. Doctors cannot reverse that order, but they can still fight over the terms.