In November 2020, Bisi Bennett went into labor at seven months pregnant. Halfway to AdventHealth hospital, with her husband, Chris, behind the wheel, she gave birth to their son Dorian in their Mitsubishi Outlander. He didn’t breathe. At the hospital, before they wheeled Bisi away from her newborn, she heard four magical words: “We’ve got a pulse.” Dorian spent 56 days in the neonatal intensive care unit getting respiratory support, surgery, cardiology, and nutritional care, and thrived. The doctors and nurses had, by any measure, performed a small miracle. It’s hard to put a price on a miracle. But technology has increasingly taken on the challenge. Bisi, who works in the insurance industry, had done everything right, financially speaking. She’d attended an in-network hospital close to home. She watched their deductible carefully. Then the bill came. More than $550,000. The hospital’s proposed installment plan: $45,843 a month for a year. Somewhere between Dorian’s discharge and the family’s mailbox, the hospital’s system had swallowed the electronical medical records and automated, algorithmically driven computers spat out numbers that took on a life of their own. Those numbers crashed against similar systems, this time on the side of insurance companies, which fired their own numbers right back. The systems locked in battle and the family—with no one to defend them—were collateral damage. It was only after a reporter poked around (as reported by Kaiser Health News), that actual people reviewed the bills and intervened as peacemakers in the cyberwar, saving the family half a million dollars. Bisi Bennett’s experience is the story of a financial side effect — the downstream damage that medical care increasingly causes not through clinical error but through the increasingly more sophisticated AI-powered engines behind payment and billing. Last December, in a widely read New York Times op-ed, Yale economist Zack Cooper explained, with precision, why American health care costs so much: Hospital prices have grown faster than almost any other sector of the economy over two decades, driven by mergers, consolidation, and the effective elimination of competition in local markets. He is right, and his research is damning. But there is a layer beneath the price problem that Cooper’s framing doesn’t fully illuminate. The question of why individual bills are so large is separate from the question of how they get so large in the first place — and the answer to the second question involves something stranger and newer than market consolidation. It involves an arms race between artificial intelligence systems, fought in the dark, on your behalf, without your knowledge, over a document you will never fully understand. Here is how a hospital bill is actually born. After a patient is discharged, the hospital produces what’s called a “superbill” — a detailed ledger of every service, procedure, and medication. That superbill gets translated into a diagnosis-related group, or DRG, a standardized code that determines the base payment an insurer will make. But DRGs are only the starting point. Hospitals can increase payment substantially by identifying what are called major complications and comorbidities, or MCCs, and the slightly lesser complications and comorbidities, or CCs — secondary diagnoses that, if documented in the medical record, justify billing the patient as sicker, and therefore more expensive to treat. Think of it this way. A hotel charges a base room rate. But the final bill depends on whether you ordered room service, used the minibar, or incurred late checkout fees. In health care, the minibar is a secondary diagnosis — and the financial incentive to find one, document one, or code one more aggressively than the clinical reality warrants is enormous. For decades, this coding was done by human specialists. Not anymore. Hospitals are now deploying AI-powered “ambient listening” software that records physician-patient conversations and automatically populates the electronic health record with diagnoses. These systems, often marketed to reduce physician burnout, have a secondary effect: They identify and document secondary diagnoses that a busy clinician might not have noted in the chart. Conditions that were once coded occasionally now appear routinely. And because hospitals get paid more when patients appear sicker, the financial gravity emanates in one direction. The data are not subtle. Blue Cross Blue Shield recently analyzed tens of thousands of maternity admissions nationwide and found a striking pattern: At hospitals with the fastest adoption of AI coding tools, the proportion of patients coded with “acute posthemorrhagic anemia,” or serious post-delivery bleeding, jumped from roughly 4% to more than 12% between 2022 and early 2025. At hospitals without those tools, the rate barely moved. Yet the increase in diagnoses was not matched by a corresponding increase in treatment; rates of blood transfusion, which you’d expect if more women were truly hemorrhaging severely, remained flat. When auditors examined the records at one hospital where the spike was most dramatic, fewer than 20% of the coded cases met clinical criteria for the diagnosis. BCBS found that AI-enabled coding added an estimated $22 million to maternity costs alone at the studied hospitals in a single year, with total nationwide effects potentially reaching $2.3 billion. Maternal hemorrhage is just one data point. In Massachusetts, hospitalizations for septicemia — the body’s catastrophic response to infection, a condition that is genuinely life-threatening and expensive to treat — have more than tripled since 2010, reaching more than 42,000 cases in the year ending last September. Sepsis is now the third-leading cause of hospitalization in the state. But researchers, insurers, and even some hospital clinicians are skeptical that Massachusetts has actually become that much sicker. The Massachusetts Health Policy Commission has documented what the data suggest: It is the highest-severity cases, the ones that fetch the largest reimbursements, that are growing fastest, without corresponding increases in length of stay or intensive care utilization. “There’s no way that’s reflecting the actual increase in the health status of the state’s population,” one state official told Modern Healthcare. “We’ve documented before that this is really about coding behavior.” Sepsis, it turns out, is one of the most lucrative secondary diagnoses in the Medicare billing system. In fiscal year 2019, the single most frequently billed Medicare DRG in the entire country was severe sepsis with major complications — a $7.4 billion line item, for 581,000 patients, with average hospital stays that were suspiciously short for patients described as critically ill. The government’s own watchdogs noted the contradiction plainly: The diagnoses suggested very sick people, but the length of stay suggested something else. The most consequential deployment of AI in American health care today has nothing to do with diagnosing cancer, predicting heart attacks, or accelerating drug discovery. It is the automated pursuit of the most lucrative billing codes for every patient who walks through a hospital door. On one side of this contest sit hospitals and physician groups, deploying “revenue cycle management” systems — AI platforms that scour medical records to optimize coding, anticipate payer objections, and fight denied claims. The revenue cycle management industry was valued at roughly $65 billion in 2025. On the other side sit insurers, deploying “program integrity” AI to detect patterns suggesting upcoding, flag suspicious claims, and deny reimbursements automatically. Both sides have named their work something that sounds clinical and responsible. Both sides are, in effect, training algorithms to fight over the same patient record. The waste this generates is staggering. Health systems collectively spend more than $140 billion annually on revenue cycle management alone. Administrative expenditures consume more than 40% of total hospital spending. Hospitals now conduct nearly half a million post-claim inpatient reviews per year for a single large insurer that often denies claims using their own versions of AI — each one a small battle in a war neither side can win decisively, because the other side adapts. The AI doesn’t rest. It learns. Then the other AI learns too. Meanwhile, the patient — the person who gave birth in a car, or had sepsis, or needed 56 days in a NICU — is not at the table. There is a famous diagram of Napoleon’s march to Moscow, popularized by the designer Edward Tufte. It shows the Grande Armée in June 1812, some 422,000 soldiers wide, crossing the Niemen River. By the time they retreated through that same border in December, the band of survivors had thinned to fewer than 10,000. They were not defeated in a single battle. They were worn down mile by mile, cold by cold, delay by delay until the army that existed at the end bore no resemblance to the one that had set out. American patients follow a similar arc. The encounter with a physician is usually the widest point of the band: the moment of greatest trust, highest attention, most direct human connection. But after that encounter ends, the financial engine engages and every step generates friction. A test requires up-front payment. A referral hits an insurance wall. A prescription turns out to be on the wrong formulary tier. A claim gets denied and appealed. A bill arrives, then another, then a collection notice. Each individual obstacle is, in isolation, manageable. But they compound. And as they do, patients thin out. They skip follow-up visits. They don’t fill prescriptions. They decide, quietly, that the system is not worth fighting. When they reappear — if they reappear — their conditions have often progressed. The financial side effects of care have become clinical ones. The payers are not entirely wrong to resist upcoding. The hospitals are not entirely wrong to insist they should be paid for every legitimate diagnosis. But as each side deploys more sophisticated tools to protect its position, the friction they generate falls not on each other but on the people caught between them. A different architecture is possible — one that treats the problem not as a permanent arms race to be managed but as a structural failure to be solved. One proposal currently being explored by some organizations (including mine) would fuse the two competing AI systems into a single adjudication engine — ingesting clinical records and billing data simultaneously, at the point of discharge, and issuing a single consensus payment decision in near real time. The idea is essentially a ceasefire proposal. Rather than having hospital AI and insurer AI fight over the same record for months, a unified system would evaluate both the provider’s right to accurate reimbursement and the payer’s obligation to appropriate payment at the same moment. Most claims would pass through immediately, giving hospitals faster and more predictable cash flow. Claims with genuine discrepancies would be flagged for immediate human review, with documentation already in hand. This is not a panacea. It requires unprecedented data-sharing between institutions that have spent decades treating each other as adversaries. It demands governance frameworks that don’t yet exist at scale. And it leaves open harder questions about the underlying price levels that Cooper and others have documented — questions that will require legislative action, not just technological creativity. But it points toward something important: the possibility of a system that settles the question of what care costs before the patient ever sees a bill, rather than after months of algorithmic trench warfare. The last thing the emergency crew told Bisi Bennett, as they wheeled her away from her newborn son, was that he had a pulse. That was the moment that mattered. Everything that followed — the 56 days in the NICU, the extraordinary clinical care, the small miracle of Dorian going home — was what the system was built to deliver. The half-million-dollar bill was what the system delivered as well. Not because anyone acted in bad faith. Because the engine to pay for heroism has become heroic in its own right — endlessly inventive, strategically brilliant, and almost entirely disconnected from the patient whose name appears at the top of the page. We can fix this. But not by pretending the arms race doesn’t exist, and not by letting it run unchecked. The first step is being honest about what the war is actually costing us. Darshak Sanghavi, M.D., a pediatric cardiologist and former senior federal health official, is chief medical officer of Machinify.
