The Blue Cross Blue Shield Association (BCBSA) says AI-enabled medical coding tools helped add an estimated $942 million in hospital spending for Blue Cross and Blue Shield plans over two years, 2024 and 2025, measured against 2023 coding levels. The analysis, which the association released on September 24 and which drew wider coverage in the following days, argues that hospital claims were increasingly classified as medically complex while the treatment patients actually received did not change to match. The American Hospital Association (AHA) disputes the reading. Here is what was reported, how the mechanism works, and what remains unproven.
What the BCBSA analysis reports
According to BCBSA's release and the press coverage of it (CNBC, Fierce Healthcare and others), the association examined claims from its member plans, which cover roughly one in three Americans, for the period from the first quarter of 2023 through the fourth quarter of 2025. The headline findings as reported:
- Hospital claims classified as medically complex rose from about 37% of inpatient stays at the start of 2023 to about 40% at the end of 2025.
- The association estimated that shift produced about $942 million in additional spending across its plans, with more than 55,000 additional stays landing in higher-severity categories.
- About 70% of the increase, roughly $650 million (reported as $653 million in some coverage), came from secondary diagnoses that moved a claim into a higher-paying severity group.
- BCBSA said more than 60% of hospital systems now use AI-enabled tools that scan lab results and electronic records for secondary diagnoses.
One detailed write-up describes the underlying white paper as a deep dive on major bowel procedures (MS-DRG 329 to 331), using de-identified claims. That scope matters: the per-procedure severity figures come from one procedure family, while the $942 million is presented as a broader estimate. I could not retrieve the full BCBSA document directly, so the exact extrapolation method behind the total is something to check against the primary release.
Where BCBSA says the $942 million came from
Estimated additional hospital spending for Blue Cross and Blue Shield plans, in millions of dollars.
Source: Blue Cross Blue Shield Association analysis, September 2026, as reported by CNBC and others.
The chart above shows the only breakdown the coverage supports. The roughly $289 million remainder is simple subtraction (942 minus 653), and the coverage I reviewed did not itemize it, so it should not be read as a defined category.
How coding intensity works
Hospitals are generally paid for inpatient stays through diagnosis-related groups. For many conditions and procedures, the same base group splits into three tiers: no complication or comorbidity, a complication or comorbidity (CC), or a major one (MCC). Each step up pays more. The tier is determined by the secondary diagnoses documented in the record, such as anemia, malnutrition or kidney disease.
That structure is the reason coding intensity is a recurring policy topic. A patient with the same operation can move up a tier if an additional qualifying diagnosis is documented and coded. If the condition was real but previously undocumented, the higher payment reflects care that was always needed. If the diagnosis rests on thin evidence, it is what payers call upcoding. The claim itself cannot tell you which.
There is a clear disconnect between coding and treatment.
BCBSA, as quoted by Breitbart's summary of the analysis
BCBSA's case rests on exactly that gap. As reported, hospitals with the biggest severity increases did not show matching increases in resource use. One example cited was anemia: diagnoses rose while transfusion rates did not follow. One summary reported ICU use of 11.5% in the top quarter of hospitals by severity growth versus 13.2% elsewhere, though that comes from a single secondary write-up and should be confirmed against the source.
Where AI enters the workflow
Two product categories feed this debate. Ambient scribes record the clinician-patient conversation and draft the note, which can surface details that a rushed manual note would omit. We covered the public-sector version of that market in our look at the VA's $775M ambient AI contract. Computer-assisted and autonomous coding tools read the chart, including lab values and notes, and suggest or assign codes, which is where BCBSA focuses.
The association's specific concern, according to CNBC and other coverage, is diagnoses derived from single laboratory values flagged by software rather than from a clinician's documented judgment. Brown University economist Christopher Whaley framed AI as accelerating existing billing incentives rather than creating new ones. The analysis does not name specific hospitals or AI vendors, and BCBSA acknowledged it cannot independently verify whether individual diagnoses were legitimate.
The hospital counterargument
The AHA's position is that the claim AI has driven up coding intensity is not supported by the evidence. Its argument, as reported:
- Patients are older and more clinically complex. The AHA has cited a roughly 5% rise in case-mix index from 2019 to 2024, and attributed about 19% of expense growth to greater patient complexity.
- Milder cases have shifted to outpatient settings, leaving a sicker mix of inpatients.
- AI helps clinicians document and capture conditions that were always present, which is the accuracy argument for these tools.
- Insurers use their own automated downcoding and denial practices.
Both readings are consistent with the same raw trend. A rising share of high-severity claims fits a sicker population, better capture, aggressive coding, or a mix. Distinguishing them needs clinical records, which claims data lacks. Coverage also notes a possible soft spot in the adoption figure: one analysis traces the 60% number to a Healthcare Financial Management Association survey of broader revenue-cycle automation, not specifically hospital coding AI.
Payer AI versus provider AI
This dispute sits inside a wider arms race. Hospitals deploy software to capture every supportable diagnosis and defend claims; insurers deploy software to detect pattern shifts, deny or downcode claims, and now publish analyses like this one. Each side points to the other's automation as the source of friction. BCBSA's own framing connects the finding to costs, with association representative Luke Chalker saying such costs can eventually show up in higher premiums and out-of-pocket spending.
The history is familiar. Coding intensity predates AI: it was a documented effect after the introduction of severity-adjusted payment groups, and it has long been a focus in Medicare Advantage risk adjustment. What is new is the speed and consistency at which software can scan a chart, which is why the question is whether AI changes the scale rather than the nature of the behavior.
The regulatory angle
CMS has not, in the coverage reviewed, announced action tied to this analysis. One report quotes CMS Administrator Dr. Mehmet Oz as saying that in the short term AI is going to be inflationary because it will turbocharge the ability of current billing systems to work more effectively; that quote comes from a single outlet and is worth verifying. Existing levers include payment audits, documentation requirements and adjustments to severity weights. Whether any of them is applied to AI-assisted coding specifically is an open question.
What to watch
- The primary data. Whether BCBSA publishes methodology for the $942 million extrapolation beyond the bowel-procedure analysis, and whether outside researchers can replicate it.
- Clinical validation. Auditing whether AI-suggested secondary diagnoses were supported by treatment and clinician documentation, the evidence both sides lack today.
- Vendor accountability. Whether coding-tool makers face disclosure or audit expectations, given the analysis names none.
- Negotiations. Payer-provider contracts may increasingly include clauses on AI-assisted coding and clinical validation.
For teams evaluating AI in documentation workflows, the lesson is auditability: keep the human judgment and the evidence trail attached to each code. Platforms that give access to several models, such as Metir, make it easier to compare outputs, though no tool resolves the underlying incentive question.
The $942 million is an estimate from a party with a financial stake, disputed by another party with a financial stake. The numbers are real data points; the causal story is an inference both sides will keep testing.
Sources:
- BCBSA: Analysis of how AI coding tools affect healthcare costs
- CNBC: Health insurer points finger at AI as nearly $1 billion in questionable hospital charges appear
- Fierce Healthcare: Hospitals' use of AI coding tools cost BCBSA plans $942M more for similar care
- Dallas Express: Blue Cross says AI-enabled hospital coding added $942 million in costs
- Breitbart: Blue Cross Blue Shield, AI-assisted medical coding added $1 billion in hospital costs
- XenoSpectrum: BCBSA AI hospital coding $942M
- BERI: Blue Cross blames $942M on AI coding that treatment never followed
Image credits
- Hero: "Cardiac surgery operating room" by Pfree2014, Wikimedia Commons, CC BY-SA 4.0. Generic illustration.
- In-body: "Operating theatre" by Piotr Bodzek, MD, Wikimedia Commons, CC BY-SA 3.0. Generic illustration.