The short answer
The first large payer dataset on AI in medical coding says it raises bills, not care. BCBSA reports that the share of inpatient stays billed as medically complex went from 37% to 40% between early 2023 and late 2025, mostly because more secondary diagnoses were recorded. It puts the cost to its plans at about $942 million over 2024 and 2025.
Whether that is fraud, catch-up documentation or sicker patients will be argued for months. The practical result is already clear: payers will now audit AI-assisted codes harder, and insurers are buying their own AI to push back. Teams building HealthTech software such as ambient scribes, clinical documentation tools or coding assistants should expect to prove that each code maps to documented care.
What did Blue Cross find?
BCBSA, the national association of the Blue Cross and Blue Shield companies, compared inpatient claims from 2024 and 2025 with a 2023 baseline. Across that period, hospitals described more patients as having complex conditions. The association links the rise to hospitals’ adoption of AI tools that read the medical record and surface diagnoses for coders and clinicians to document.
“We find no evidence of a corresponding change in care,” Luke Chalker, BCBSA’s senior vice president of product and data science and a co-author of the analysis, told reporters. He told The New York Times that “coding is changing because of this technology.” Speaking to TechCrunch, he described the contest between hospital and insurer software as one-sided in the hospitals’ favor.
How do AI coding tools change a hospital bill?
US inpatient stays are paid by diagnosis-related group, or DRG. The same main diagnosis can pay very different amounts depending on whether the record also documents complications or comorbidities. A secondary condition such as anemia or low sodium, once written into the chart and coded, can move a stay into a higher-paying group.
AI documentation and coding tools are good at exactly this: they scan notes, labs and history, then prompt the clinician or coder to record conditions that were present but not written down. That can be legitimate. It can also inflate bills when a condition is recorded but did not affect treatment. BCBSA’s point is that secondary diagnoses account for about 70% of the added cost while the care itself looks the same.
Is it upcoding or better documentation?
Hospitals reject the framing. The American Hospital Association said in August that higher coding intensity can reflect older and sicker patients and more complete records. It cites an AHA and Vizient analysis showing hospital case-mix index up about 5% between 2019 and 2024, and points to insurers’ own coding practices, including a 2025 MedPAC finding that upcoding contributed to about $40 billion in overpayments to Medicare Advantage plans.
Both sides are now using AI. As The New York Times describes it, insurers can answer with software that downcodes claims or recalibrates risk scores, and hospitals can answer with better documentation tools. Health policy researchers warn that this arms race could consume the efficiency gains AI was supposed to bring, with employers and patients paying through premiums. Abridge founder Shiv Rao told TechCrunch the competing systems could lead to a dystopian outcome, though he argued AI might also reduce friction and cost.
What it means for US & EU software teams
First, explainability is now a sales requirement. A hospital buying a coding or documentation assistant will be asked by its own compliance team, and eventually by payers, why each code was added. A model that only outputs a list of codes is hard to defend. A system that shows the sentence, lab value or order supporting each diagnosis, and records who accepted it, is much easier to buy.
Second, audit trails have to be designed in, not exported later. Store the model version, the source evidence, the suggestion, the human decision and the timestamp for every coded item. In the US, that record is what stands between a vendor and False Claims Act exposure if a payer or regulator alleges systematic upcoding. Because it contains protected health information, it also has to meet the access, encryption and logging rules of HIPAA-compliant software development.
Third, monitor for drift at the portfolio level. BCBSA’s method is simple: compare coding patterns to a baseline and look for changes that treatment does not explain. Product teams can run the same check on their own customers. A sudden jump in complications per stay after a model update is a signal to review before a payer finds it. For EU teams, the same logic applies in DRG-based systems, and AI that feeds clinical or reimbursement decisions also has to meet GDPR health-data rules and, where it qualifies as high-risk, EU AI Act logging and oversight duties.
What to do now
- Link every suggestion to evidence. Show the exact note text, lab or order behind each proposed diagnosis, and do not suggest codes without it.
- Keep a human in the loop. Require a coder or clinician to accept each suggestion, and record who did it and when.
- Log the full chain. Model version, input, output, decision and final code, retained for the payer audit window.
- Track coding intensity. Monitor complication and comorbidity rates per stay against a pre-deployment baseline, and alert on unexplained jumps.
- Review claims logic before model updates. Treat a model change that shifts coded severity like a pricing change, with sign-off from compliance.
Frequently asked questions
What did the Blue Cross Blue Shield Association find?
In an analysis released on September 24, 2026, BCBSA said the share of inpatient stays billed as medically complex rose from 37% at the start of 2023 to 40% by the end of 2025. It estimates the higher coding intensity cost its member plans about $942 million over two years, and says it found no evidence that the care delivered had changed.
How do AI coding tools raise a hospital bill?
Inpatient stays are paid by diagnosis-related group. Adding documented secondary diagnoses, such as anemia or low sodium, can move a stay into a higher-paying group. BCBSA attributes $653 million of the $942 million to secondary diagnoses, roughly $11,000 per excess complex case, and links the shift to AI tools that capture more conditions from the record.
Is this upcoding or better documentation?
That is the dispute. BCBSA says coding changed without a matching change in treatment. The American Hospital Association has argued that coding intensity also reflects older, sicker patients and more accurate documentation, citing a roughly 5% rise in case-mix index from 2019 to 2024, and points to insurer upcoding in Medicare Advantage.
What does this mean for companies building clinical AI?
Expect payers and auditors to question every code your product suggests. Each code should link to the exact evidence in the note, show who accepted it, and be logged. Vendors that can prove a code is supported by documented care will be easier to buy and easier to defend in an audit.
Does this affect teams outside the US?
The numbers are US claims data, but the pattern applies wherever payment depends on coded diagnoses, including DRG-based systems in the EU. In the EU, AI that influences clinical or reimbursement decisions also meets GDPR rules on health data and, in some cases, EU AI Act obligations for high-risk systems.
Sources
TechCrunch — Insurers claim AI is already increasing healthcare costs
The New York Times — reporting on the BCBSA analysis of AI and hospital billing, September 24, 2026
Fierce Healthcare — Hospitals’ use of AI coding tools cost BCBSA plans $942M more for similar care: analysis
Blue Cross Blue Shield Association — AI coding intensity white paper (PDF)
Medical Daily — Blue Cross ties $942 million in added hospital costs to AI coding, but hospitals say patients are sicker