Blue Cross Study Finds AI Documentation Tools Added $942 Million to Insurer Bills

The AI tools hospitals bought to cut paperwork are now at the center of a nearly $1 billion billing dispute. A claims analysis released Thursday by the Blue Cross Blue Shield Association estimates that heavier use of AI-assisted coding and documentation cost its member plans $942 million more for inpatient care in 2024 and 2025 than 2023 billing patterns would have. The association says there is no sign patients got more treatment for that money.

The study covered inpatient hospital claims across BCBSA's 31 independent Blue companies, which insure more than 100 million people. About $653 million of the extra spending came from secondary diagnoses, meaning conditions such as anemia and low sodium that get recorded alongside the main reason for a hospital stay. When those diagnoses are added to a claim, they can move it into a higher-severity, higher-paying diagnosis-related group, or DRG. BCBSA puts the cost at about $11,000 per excess complex case.

Diagnoses up, treatment flat

BCBSA says the share of Blue-billed inpatient stays classified as medically complex rose from 37% in early 2023 to 40% by the end of 2025. About 70% of that increase came from more than 55,000 additional cases in which a secondary diagnosis moved the claim into a higher tier.

Major bowel surgery shows the pattern most clearly. Among these patients, recorded diagnoses of partial intestinal blockage rose 55% between the first quarter of 2023 and the fourth quarter of 2025, and diagnoses of excess acid in the blood rose 33%. Over the same period, claims coded at the highest complexity level for major bowel procedures more than doubled, from 10.2% to 22.7%. That DRG alone accounted for nearly $61 million of the added costs, according to Fierce Healthcare's account of the white paper. The treatments those diagnoses would normally lead to did not rise to match.

"If patients are truly sicker, we'd expect to see more treatment," Luke Chalker, BCBSA's senior vice president of product and data science and one of the report's authors, told Reuters. "The disconnect between diagnoses and treatment suggests that AI is identifying more billable conditions, not sicker patients."

Anemia is the association's clearest example. Posthemorrhagic anemia, which the white paper calls a common "bump code," was diagnosed in 13.7% of cases at the top quarter of hospitals, compared with 9.9% everywhere else. Yet patients diagnosed at those high-coding hospitals received transfusions less often, 16.9% versus 19.3%. More broadly, the hospitals coding the most complex cases had similar or lower rates of ICU use, reoperation, transfusion and length of stay than their peers.

"The question that is worth asking is [with] two similarly situated hospitals, treating similar patients, why would one hospital diverge?" Dr. Razia Hashmi, BCBSA's vice president of clinical affairs, told reporters at a briefing. "There may be an element of correct coding there, but the likelihood that this is technology-enabled upcoding is higher, in my view."

BCBSA ties the increase to "systemic" adoption of AI revenue-cycle software. That includes tools that comb existing medical records for conditions nobody coded and ambient scribes that listen to doctor-patient conversations and draft clinical notes. The association cited a June survey in which more than 63% of healthcare organizations said they use AI in revenue-cycle work. Chalker said the $942 million counts only cases where BCBSA found no change in the care delivered. Claims where hospitals recorded additional treatment were left out. "That's the stuff that hospitals should bill for, and that's the stuff we should pay for," he said.

The hospitals' side

Hospitals reject that reading. In August the American Hospital Association argued that rising coding intensity can reflect an older, sicker inpatient population, with simpler care moving to outpatient settings, as well as more accurate documentation of conditions patients already had. The AHA said hospital case-mix severity rose about 5% from 2019 through 2024. Health systems also describe AI coding as a response to what they see as aggressive denials and payment delays from insurers. Some openly say it pays off: Michigan-based McLaren Health Care has said AI-enhanced documentation brings in about $1 million a month in additional revenue.

BCBSA admits a real limitation. The study uses claims data rather than patient charts, and charts are the direct way to check whether patients were actually sicker. Chalker said Blue plans that do have chart access through provider relationships have seen the same effect.

Why It Matters

Ambient scribes were sold on a simple promise: less typing for doctors and less burnout. Abridge, Microsoft's Nuance DAX and a growing list of competitors have spread quickly through health systems on that pitch. This report points to what happens when a tool that records more of what a doctor says is paired with a payment system that pays more for every documented condition. Some of that extra documentation is legitimate. Clinicians have long under-coded real conditions. But BCBSA's inverse relationship, where the hospitals coding the most disease deliver the same or less treatment, is the sort of pattern that has drawn fraud scrutiny in Medicare Advantage risk adjustment.

It is also an escalation in an AI arms race. Hospitals use models to find every billable code, and insurers use models to flag, deny and audit them. Either way, employers and patients pay through premiums. BCBSA's March analysis had already tied up to $2.3 billion in inpatient and outpatient spending to AI-enabled coding. This new report puts a specific per-case cost on the inpatient side.

What to Watch

BCBSA says it will publish more analyses covering outpatient care and other high-volume DRGs. Expect the Blue plans to use these reports to justify tighter payment policies, such as automatic downcoding, clinical validation audits, or contract terms that set rules for hospitals using AI documentation tools. Watch whether CMS or state regulators start asking scribe and coding vendors to show their tools are not pushing diagnoses without clinical support. Until someone produces chart-level data, both sides will keep arguing over the same claims.