This is a Physician in the Loop special topic. Proof should set the price Medicare has renamed the diagnostic software it pays hospitals for and parked it in temporary payment groups, an interim step. The next step should pay for software that shows patients do better. Medicare already pays hospitals for diagnostic software. From January it will do so under a name of its own. Its final outpatient rule for 2027, issued on Nov. 4 beside a cut in physicians' pay rate, renames programs that analyze a patient's data and return a diagnosis, a risk score or a treatment recommendation. Software as a Service becomes Software as a Medical Service. For now these services go into payment groups Medicare keeps for new arrivals, a move the Centers for Medicare and Medicaid Services, or CMS, calls "an interim step toward a broader payment policy." That broader policy is not yet written, and it matters. Hospitals buy the software that earns its keep, and Medicare's rules decide what earns. Payment is among the strongest levers over which AI reaches patients. When the interim ends, Medicare should pay for software that shows its patients do better, not merely software that has been cleared and sells well. The new category's rules should say so. CMS has sound reasons to go slowly. In its July proposal it said its outpatient prices are built around supplies, equipment and staff time, which say little about what an algorithm is worth, and some vendors charge by subscription, by license or by the click. The proposal named 36 billing codes, from AI reading of retinal photographs to estimates of blood flow in the heart's arteries from a CT scan. The rule parks such services in groups that pay for newcomers until CMS has gathered enough claims data to place them for good. Under the outpatient system Medicare pays the hospital, not the vendor. The second change reaches patients directly. A small number of analyses leave the lab fee schedule, under which Medicare patients generally owe nothing. They work only on data a laboratory has already produced. One of the ten in the proposal was a secondary computer interpretation of a genomic sequence. The outpatient system, unlike the lab fee schedule, charges patients a share, so from January some will owe a copay for them. CMS says the analyses are substantially similar to software it already pays hospitals for, and it set their new rates in line with the old. The Association for the Advancement of Blood and Biotherapies had warned that the move would subject the analyses to budget neutrality as well as cost-sharing, "potentially shifting costs to beneficiaries and providers." Nothing in CMS's summary of its software policy, or in its criteria for the new-technology groups, asks whether patients do better. To enter those groups a service must be new, must not be eligible for pass-through payment and must be "reasonable and necessary," the ordinary test for coverage. Under the usual rules for those groups, the permanent price would come from claims, which pile up fastest for what sells. Yet CMS knows how to ask. In July it finalized the repeal of a shortcut that let devices with the Food and Drug Administration's breakthrough designation collect extra inpatient and outpatient payments without showing that they substantially improve diagnosis or treatment. From the 2028 round of applications, after a transition, they must show it again. An agency that has just ruled that an FDA designation cannot stand in for proof of benefit should not let a sales record stand in for it either. Evidence of that kind can be had. Take the estimate of blood flow in the heart's arteries from a CT scan. In a randomized trial of 2,103 patients with stable symptoms of possible coronary disease, a strategy that used it sharply cut the number of patients sent for a catheter test, a tube threaded to the heart, only to find no significant blockage. With usual testing that was about one patient in ten. With the new strategy it was one in forty. Those are people spared a needless procedure. The trial was sponsored by HeartFlow and bundled the software with a risk score and the scan, and its authors concede that "the separate effects of each choice cannot be determined." Heart attacks, more than twice as common with the new strategy, were too few to draw a conclusion. Still, it is evidence a payer can price. There is a serious case for paying first. Diagnostic software helps a patient only through what a doctor does with its answer, so trials that follow patients are slow and costly. Why, the argument goes, should an algorithm face a test that routine lab work is spared? Demand such trials and only the largest companies will afford them, while patients wait for tools that may well work. Grant all of it. It argues for paying in stages, a base rate through the interim and more once trials in patients show a benefit, so that Medicare's money pays for the evidence instead of standing in for it. CMS should build the broader policy on what happened to patients. A code backed by trials showing fewer missed diagnoses, needless procedures or deaths should earn more than one backed only by accuracy on stored images, and the evidence behind each code should be posted beside its price. The outpatient rule was issued as final with a comment period, so medical societies should make that case in writing before the next rule is drafted. Hospitals should ask vendors for the same evidence before they sign, and physicians should warn patients that some analyses will carry a copay from January. The nearly 1,400 critical access hospitals are paid on their own costs, not these rates, so a price that rewards evidence would miss the smallest rural hospitals. A coverage rule would reach them. CMS called its decision an interim step, and it is. For now the agency can still choose what its money rewards. Once the claims arrive, they will choose for it, and a claim records what was billed, not how the patient fared. This special topic was read by an AI voice. Its sources are linked at physicianintheloop.org.