What the analysis says
The Blue Cross Blue Shield Association published a claims analysis on 24 September estimating that hospitals’ growing use of AI coding tools added about $942 million to what its member plans paid over two years, without a matching change in the care delivered, Fierce Healthcare reported.
The headline movement is small and expensive. The share of inpatient cases billed to Blue plan members as medically complex rose from 37 per cent at the start of 2023 to 40 per cent at the end of 2025. About 70 per cent of that increase came from more than 55,000 additional cases in which secondary diagnoses pushed a claim into a higher-severity, higher-reimbursement diagnosis-related group. That slice alone accounts for $653 million, or roughly $11,000 per extra complex case.
These are the association’s own figures about claims its members paid, and they are an estimate, not an audit.
The argument, and its limits
BCBSA’s case is that coding changed and treatment did not. “Critically, what we found is underneath all of that data no change in corresponding care for a more complex patient,” Luke Chalker, the association’s senior vice president of product and data science and one of the authors, told reporters. “We now see that coding has materially changed. We see that. But we find no evidence of a corresponding change in care, and that’s because the reimbursement mechanisms that exist allow this.”
The supporting evidence is comparative. Looking at the top quartile of hospitals by complex cases, the association found similar or lower treatment intensity — ICU use, transfusion, reoperation, median length of stay — than peers. On posthemorrhagic anemia, which the paper calls “a common bump code”, the hospitals diagnosing it most often (13.7 per cent versus 9.9 per cent) transfused those patients less often (16.9 per cent versus 19.3 per cent).
BCBSA acknowledges the limit: the work rests on claims, not clinical documentation, which would be the direct test of whether those patients really were sicker.
The other side
Hospitals have argued that AI revenue-cycle tools help them code more accurately and efficiently under increasing payer scrutiny, and that rising coding intensity can also reflect an older, sicker population and better documentation of conditions that were always there. Chalker rejected the framing that the tools are simply securing payments long owed, arguing that costs rising with no added value for anyone downstream is a change for the worse.
TechCrunch, reporting the analysis on 26 September, quotes Chalker describing the position insurers find themselves in more bluntly: “It’s not a war. It’s a completely one-sided blood bath.” It also quotes Abridge founder Dr Shiv Rao warning of “a horrible dystopic future nobody wants to live in” with “bots fighting bots, agents fighting agents”.
Why this is an AI story
Because it is one of the first attempts to put a number on a second-order effect. The tools in question are not clinical models. They read patient records and lab results and find anything that can legitimately be coded — a productivity gain, precisely applied, inside a payment system that rewards severity.
BCBSA cites a June survey in which more than 63 per cent of healthcare organisations said they use AI in revenue cycle workflows. The disagreement is not about whether the tools work. It is about what “working” means when both sides of a transaction automate.
What to watch
BCBSA says this white paper covers inpatient services only, and that it will publish further analyses on outpatient care and other diagnosis-related groups. The stronger test would be a version built on clinical charts rather than claims, which Chalker said Blue plans with provider relationships have begun to do.
