Special Topic: Nobody wins the bot war

Show Notes

Hospitals and insurers now fight over bills with software, and patients pay. Every algorithm that moves a payment should show a clinician its evidence and be judged by how often it is right.

Read the special topic and its sources: https://physicianintheloop.org/specials/nobody-wins-the-bot-war/

The voice on this episode is generated by AI. It reads a special topic from Physician in the Loop as written; every fact in it links to its source.

What is Physician in the Loop?

How technological innovation is changing medicine for patients and the people who practice on the front lines.

This is a Physician in the Loop special topic.

Nobody wins the bot war

Hospitals and insurers now fight over bills with software, and patients pay. Every algorithm that moves a payment should show a clinician its evidence and be judged by how often it is right.

The American Hospital Association named the right test for billing software on Oct. 5. "The appropriate test for whether a patient has been correctly coded is not whether reimbursement increased; it is whether the patient's medical record supports the diagnosis," the association wrote. It was answering the Blue Cross Blue Shield Association, which estimates that hospitals' AI coding tools cost its member plans $942 million over two years for care that, it says, did not change. The consulting firm PwC told insurers much the same in June: "The goal should be more accurate payments, not more denials of claims."

Both sides now field software that is good at the fight. Hospitals use it to record what PwC calls "greater specificity and reimbursable severity" and to write appeals, and Kaiser Permanente has seen AI-written appeals raise its recovery rates by about 15%. Some insurers, the hospitals say, use automated tools to lower and partly deny claims. The chief revenue officer of New York City's public hospitals calls it a fight in which "nobody's going to win." She is right, and patients are paying for it. Every algorithm that moves a payment, up or down, should have to show a clinician its evidence from the chart and be audited for how often it is right, not for how much money it moves.

The insurers' case rests on their own claims data. The Blue Cross association found that the share of hospital stays billed to its plans as medically complex rose from 37% at the start of 2023 to 40% by the end of 2025. Hospitals in the top quarter for diagnosing anemia from blood loss transfused a smaller share of the patients they diagnosed than other hospitals did, a gap the association says "suggests that AI is identifying more billable conditions." It concedes that its analysis is limited by its reliance on claims rather than clinical documentation. The hospitals' association replies that the analysis does not appear to identify which claims involved AI tools, and that "the absence of an additional procedure does not establish that a diagnosis is invalid."

Doctors and hospitals complain of the reverse. Since July 1, Blue Cross Blue Shield of Illinois has paid some visits at a lower level than billed when it judges that the services do not justify that level, and Chicago-area clinicians say it is cutting claims for complex or longer visits automatically. The insurer says its review is "intended to appropriately support accurate payments" and did not comment on whether it uses AI. Appeals are one measure of insurers' errors, and they are rare. In 2024, by their own reports, Medicare Advantage insurers denied, in whole or in part, about one prior authorization request in 13; just over one denial in ten was appealed, and four in five appeals succeeded at least in part. The figures do not say how many decisions involved software, but most challenged denials did not hold up in full.

HCA Healthcare's finance chief said in September that hospitals are "behind the payers" in using AI on claims, and called the administrative cost on both sides "enormous." CommonSpirit Health's chief financial officer told investors on Oct. 2 that the system had "captured $200 million in financial value from artificial intelligence," though the published account of his remarks does not say which uses produced it. Both are claims about money. Aetna's chief medical officer says the early data on AI in health care are "largely inflationary," with "no real strong evidence that people are getting different clinical outcomes."

Patients pay for the contest in three ways. Insurers set premiums from projected growth in costs. In a PwC survey, seven in ten health plans named AI that brings hospitals and doctors more revenue among the three biggest drivers of next year's growth; that is their judgment, not a measurement. Patients who owe a percentage of the bill, or all of it until the deductible is met, can owe more when the bill is coded higher, up to their plan's out-of-pocket limit. A denial overturned on appeal is care approved in the end, perhaps after a delay, and such delays "may have negative effects on a patient's health." Patients seldom fight back: in federal marketplace plans, fewer than one denied claim in a hundred was appealed in 2024.

Regulators have begun to act, almost entirely on the insurers' side. At least seven states passed laws this year that, in various forms, bar insurers from leaving decisions to AI alone. Alabama's took effect on Oct. 1, and under a law signed by Gov. JB Pritzker, Illinois will require from 2028 that a person make or review every decision to lower a claim's billing code. Connecticut's comptroller has gone further for the state's own health plans, requiring that their AI be validated "for accuracy, consistency, and fairness" and that insurers and providers alike disclose when AI materially assists with, or recommends, a member's benefits or care.

Since January, federal rules have required Medicare Advantage, Medicaid and federal-marketplace plans to give a specific reason for each prior authorization they deny and to post figures on their decisions every year. Medicare's own pilot, which uses AI to review requests for selected services in six states, pays its contractors a share of the spending they avert and, Medicare says, leaves every recommendation to deny to licensed clinicians. It also cuts that pay when contractors miss standards for accuracy and timeliness, and withholds or recovers it when a denial is overturned on appeal.

Perhaps the hospitals' software is simply recording the truth. Hospitals say patients are older and sicker, coding vendors say their software catches conditions that hospitals used to miss, and the Blue Cross figures rest on claims rather than charts. Some insurers already cut bills by automated rule, so a hospital that set its tools aside would be paid less for the same work, and the deeper fault lies in paying for care service by service. Much of this may be right, and the rule proposed here asks no one to disarm; it asks for the hospitals' own test, whether the record supports the diagnosis. If the new codes are real, an audit against the chart will show it, and the same test would oblige insurers' software to point to the chart before it lowers a bill. Nor would paying differently end the contest. Medicare pays its private plans per patient, more for patients whose records say they are sicker, and Congress's advisers on Medicare estimate that the plans' own coding makes their enrollees' risk scores about 4% higher than those of similar patients in traditional Medicare, even after Medicare's correction.

The principle that matters most is that the evidence travels with the decision. Any software that adds a diagnosis to a bill, lowers a code or recommends a denial should cite the passages in the record it relied on, so that a clinician on either side can check the call in minutes. Medicare can require it of its private plans and its own contractors, building on the specific reasons it already demands for prior authorization denials, and states can write it into the laws they are passing. Hospitals can make it a condition of their contracts with coding vendors.

Just as important is how accuracy is measured. Appeals catch only the errors someone contests, and patients rarely appeal. The health department's inspector general showed a better way in a 2022 report: its reviewers checked a random sample of the denials that 15 of the largest Medicare Advantage insurers issued in one week of 2019 against Medicare's rules and found that more than one denied prior authorization in eight would probably have been approved in traditional Medicare. Samples like that, drawn yearly for every tool on both sides and reported as error rates in both directions, would show which software to trust. Any contract that pays a vendor a share of the money its software brings in or saves should also dock it for errors, as Medicare's pilot does. Medical staffs should ask for those error rates before a tool touches their patients' bills. For patients the test is simpler: the right decision, made quickly.

The hospitals' association has written down the right test, and insurers have heard the same advice. Applied to every bot on both sides, with the results made public, it could shorten the billing war. Until then nobody wins, and patients go on paying for the fight.

This special topic was read by an AI voice. Its sources are linked at physicianintheloop.org.