AI Coding Tools and Rising Case Complexity: What the BCBSA Analysis Shows, and What It Doesn’t

On September 24, 2026, the Blue Cross Blue Shield Association (BCBSA) published an analysis arguing that hospital use of AI coding tools has pushed up what its member plans pay. The association estimates about $942 million in added spending between 2023 and 2025. The American Hospital Association (AHA) responded in early October that the analysis does not show what it claims to show.

This is one report from one side of a payment dispute, so we read it alongside the hospital response, the Office of Inspector General’s 2026 compliance guidance and reporting from health system IT leaders. Here is where they agree, where they do not, and what an organization using AI in coding or documentation can do now.

What the BCBSA analysis found

• It used de-identified claims from Blue plans, which cover about one in three Americans, and looked at inpatient stays from early 2023 through the end of 2025.
• It estimates about $942 million in added spending. Roughly 70 percent, more than $650 million, came from secondary diagnoses that moved stays into higher-paying diagnosis related groups (DRGs).
• The share of stays billed as medically complex rose from about 37 percent to 40 percent, across more than 55,000 complex cases, according to reporting on the analysis.
• Its main example is anemia after major bowel surgery. Hospitals with the fastest-growing coding recorded more anemia diagnoses, but BCBSA reports no matching increase in transfusions.
• It points to AI tools that scan labs, notes and records to flag secondary diagnoses, and cites an outside source that more than 60 percent of hospital systems use AI coding tools. A single abnormal lab value can support a secondary diagnosis, which is why these tools find so many.
• An earlier BCBSA analysis in March 2026, which looked at maternity admissions, described a similar pattern.

What it does not show

• It uses claims, not medical records. Claims show the codes billed, not whether each diagnosis was clinically supported. By press accounts, BCBSA’s researchers acknowledge that clinical data would give a more accurate picture.
• It does not identify which claims involved AI tools. The link between AI and the rise is an inference, not a measurement. The AHA makes this point directly.
• It cannot fully separate coding from patients. The AHA says it does not account for sicker, older patients with more chronic conditions, and for lower-acuity care moving out of inpatient settings, which leaves more complex patients in the hospital.
• A flat transfusion rate is weak evidence on its own. Clinicians point out that most postoperative anemia is never transfused, so no rise in transfusions does not by itself show the diagnosis was wrong.
• The source has a stake. BCBSA represents insurers, who pay the claims. The AHA represents hospitals, who submit them. Neither side has published a chart-level review that settles it.

Insurers use AI too

Hospital leaders point out that payers use AI to review and deny claims, and that providers adopted their own tools partly in response. A 2025 AMA survey found 61 percent of physicians worried that insurers will use AI to increase prior authorization denials. That context matters, but it does not remove the question BCBSA has raised: when software suggests a diagnosis, who confirmed it and on what evidence?

What regulators are already saying

The BCBSA analysis is about commercial inpatient claims, not Medicare Advantage. But in February 2026 OIG issued its first Medicare Advantage compliance program guidance since 1999. It is voluntary, and it covers risk adjustment. Law firm summaries read it as flagging diagnoses that cannot be verified, add-only chart review, and prompts that push clinicians to add risk-adjusting diagnoses, including prompts generated by AI. Read the guidance itself before relying on a summary. The direction is clear either way: using AI is not prohibited, and the organization that submits the claim stays responsible for supporting it.

What to do now

1. Record the source of every diagnosis. Capture which suggestions came from software, which from ambient documentation and which a clinician entered, and who approved each one. Health system leaders describe this record as a foundation when dealing with insurers, and a clinician’s approval alone may not be enough if the pattern looks like upcoding.
2. Validate clinically before billing. For secondary diagnoses that change the DRG, especially those supported by a single lab value, require documented clinical support, and give coders and clinicians a way to say no.
3. Compare your own numbers before a payer does. Track case mix and secondary-diagnosis rates over time and against peers, and look at whether treatment moved with the coding. A diagnosis that rises with no change in monitoring or treatment is the pattern payers will look for.
4. Sample AI-assisted charts for audit. Review a regular sample of high-impact cases, log overrides and corrections, and use what the patterns tell you to fix prompts, rules and training.
5. Read your vendor contract. Check who owns the model’s logic, what is logged, how updates are communicated and whether any fee depends on the revenue the tool adds. Reports of vendors sharing in added revenue are the kind of detail an auditor will ask about.
6. Prepare for payer clinical validation requests. Know who responds, how fast, and with what evidence, before the first batch arrives.
7. Put AI coding under governance. Treat it as a tiered, owned, monitored use case like any other AI that touches claims.

Frequently asked questions