The anomaly showed up first in the paperwork, not the patients. Analysts at the Blue Cross Blue Shield Association were combing through two years of insurance claims when a number began to stand out: more and more patients were being documented as having complex conditions, the kind of diagnosis that lifts a claim into a higher-paying category. The care those patients received, though, looked the same as before.
What changed, the analysts concluded, was who was writing the claim. Hospitals have been adopting artificial intelligence tools to prepare insurance submissions, software that reads a patient’s chart and suggests the most profitable billing code. The association estimates the result added about $942 million in healthcare spending over two years, according to an analysis it published this week.
The finding rests on a pattern the association described as a “sharp increase in patients being documented as having complex conditions,” with “no evidence of corresponding change in care delivered.” In its words, there is a “clear disconnect between coding and treatment.”
The tools at issue are part of a fast-growing corner of the industry. Revenue-cycle vendors have spent the past few years pitching software that scans clinical notes and translates them into billing codes, promising hospitals that they will stop leaving money unclaimed. Doctors and nurses, the pitch goes, chronically under-document the severity of what they treat, and the software corrects the record. Insurers now argue the software corrects it too far, and always in the same direction.
The analysis lands in the middle of a fight that has been simmering across the industry. Hospitals and insurers have always argued over what a treatment is worth and whether it should be paid. The New York Times reported this week that the arrival of AI on both sides of the table appears to be making the friction worse, as software submits claims and other software reviews them.
Some in the industry describe the dynamic in almost martial terms. Dr. Shiv Rao, founder of the AI startup Abridge, said the technology could produce “a horrible dystopic future nobody wants to live in,” full of “bots fighting bots, agents fighting agents.” He allowed that it might also reduce tension and cut costs. Luke Chalker, a senior vice president at the association, was less diplomatic. “It’s not a war,” he said. “It’s a completely one-sided blood bath,” with insurers on the losing side.
The rise of AI coding tools has been sold to hospitals as a way to capture revenue they are otherwise leaving on the table. Billing is famously Byzantine in American medicine, with thousands of codes, and hospitals have long argued that under-coding costs them money and that medical complexity has genuinely increased as the population ages. The insurers’ report is a direct challenge to that defense, suggesting the shift is a product of software rather than of sicker patients.
There is a broader irony in the timing. Insurers have spent years marketing their own AI as a way to cut administrative waste and rein in costs, screening claims for errors and overbilling with software of their own. Now the same technology, in the hands of the other party, is being blamed for pushing spending in the wrong direction. The result is an arms race in which each side buys software to outmaneuver the other side’s software.
For patients, the stakes are concrete. Higher coding can mean higher premiums, higher out-of-pocket costs and a system in which the description of care drifts further from the care itself. The association’s report stops short of accusing hospitals of fraud, and there is no suggestion any individual provider acted illegally. But the pattern it documents, if it holds, raises questions about what the billing record will come to mean.
The dispute also sits inside a longer history. Upcoding, the practice of billing a more severe diagnosis than a patient’s condition warrants, predates AI by decades, and insurers have spent years building audits and software to catch it. What is new is the scale and the speed. A tool that rewrites thousands of charts a day can shift the entire statistical profile of a claims pool in a matter of months, which is precisely what the association’s analysts say they saw.
The $942 million is small against the trillions Americans spend on healthcare each year, but the number matters less for its size than for its direction. If AI lifts coding costs at both ends of the system, the savings the technology was supposed to deliver could vanish into the gap between what one side bills and the other side refuses to pay.
What happens next is unclear. The association has said little about whether it will press for audits or new rules. The tools themselves are still new, and hospitals are unlikely to stop using them. If the next two years resemble the last two, the $942 million figure may come to look less like an anomaly and more like a floor.
