Blue Cross Blames Hospital AI for $942 Million in Upcoding

The first big fight over what AI costs the economy is not about data centers or jobs. It is about hospital bills. On September 24, the Blue Cross Blue Shield Association published an analysis estimating that hospitals billing inpatient stays as more medically complex added about $942 million in costs for Blue Cross and Blue Shield companies across 2024 and 2025, compared with 2023. The association ties the jump to AI tools that scan charts and lab results for anything billable.

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

What the numbers say

The white paper, a three-page document titled "Hospital Coding Intensity Analysis: Major Bowel Procedures," looks at claims from Q1 2023 through Q4 2025. Over that span, the share of inpatient cases billed as complex rose from roughly 37% to about 40%. Compared with 2023 baseline rates, hospitals classified 55,158 additional cases as complex, generating $653 million in extra reimbursement at an average of $11,800 per case. That slice, about 70% of the total, came from secondary diagnoses: conditions noted alongside the main reason for the stay that push a claim into a higher-paying diagnosis-related group.

The mechanism matters. Under the DRG system, one added "complication or comorbidity" can move a surgery into a costlier tier. BCBSA says the fastest-growing of these "bump codes" are ones a machine can find in a single lab value, such as unspecified acidosis, low sodium and acute posthemorrhagic anemia, along with malnutrition and obstruction diagnoses. The paper says technologies "including ambient listening for observation codes and laboratory data mining appear to be contributing." BCBSA's release says more than 60% of hospital systems now use AI coding tools.

The anemia test

To separate sicker patients from busier coders, BCBSA looked for what it calls clinical discordance: a diagnosis on the claim without the treatment that diagnosis usually brings. It focused on major bowel procedures, DRGs 329 through 331, mostly surgeries for colon cancer or diverticular disease. From 2023 to 2025, claims at the highest complexity level rose from 20.2% to 22.7% while non-complex claims fell from 36.6% to 32.8%, adding $60.8 million in costs in that family alone.

Hospitals in the top quarter for complexity growth coded 75.6% of these surgeries as complex, versus 65.0% elsewhere. Yet their ICU use was lower, 11.5% against 13.2%, and median length of stay was the same four days. On anemia, the top-growth hospitals diagnosed it 38% more often, 13.7% versus 9.9%, while transfusing a smaller share of those anemic patients, 16.9% versus 19.3%. The paper's explanation is that AI set to flag every out-of-range lab catches mild anemias clinicians would not normally treat.

This is BCBSA's second round. In March, it published a similar study of maternity admissions that tied roughly $2.3 billion in nationwide spending, $663 million inpatient and at least $1.67 billion outpatient, to AI-enabled coding. "Something is disconnected," Dr. Razia Hashmi, BCBSA's vice president of clinical affairs, said at the time.

The hospitals' case

Hospitals have been making the counterargument for months. In a July 31 fact sheet, the American Hospital Association said hospitals "are increasingly caring for patients with higher acuity, a trend that is appropriately reflected in provider coding practices," citing an aging population, rising chronic disease and simpler care moving to outpatient settings. An AHA and Vizient analysis found hospital case-mix index rose about 5% between 2019 and 2024. The AHA also argues that AI scribes help doctors document what they found, and that clinicians carry legal obligations to code accurately.

The AHA turned the charge around, too. Its fact sheet says some insurers use "automated tools to partially deny claims," and that plans "want it both ways — a sicker enrollee population for purposes of health plan risk scores but a healthier one when it comes time to cover healthcare claims." It points to MedPAC's 2025 finding that upcoding drove about $40 billion in overpayments to Medicare Advantage plans, which insurers run.

The study also has limits worth stating plainly. It is a white paper from a trade group with a financial stake in lower hospital payments, not a peer-reviewed study, and it uses claims data rather than medical charts, which Medical Daily notes the association acknowledged would be a more direct test. The $942 million total is an estimate for the Blue system only. BCBSA says its companies cover one in three Americans.

Bots against bots

The rhetoric is escalating along with the software. Shiv Rao, founder of ambient documentation company Abridge, described where this goes as "a horrible dystopic future nobody wants to live in," with "bots fighting bots, agents fighting agents," as quoted by TechCrunch. Chalker rejected the framing of an arms race. "It's not a war. It's a completely one-sided blood bath," he said, per TechCrunch.

Both sides are telling a version of the truth. Documentation tools really do capture conditions that tired clinicians used to leave out, and under the payment rules a documented condition is a billable one. Insurers really do run automated review at scale. The problem is that a payment system built on human-speed paperwork is now being worked by software on both ends, and every extra code a model surfaces is money that someone pays.

Why it matters

This is one of the first attempts to put a dollar figure on AI's effect on a single, measurable cost line, and it will not stay a trade dispute. Since March, BCBSA has said its companies are working nationally and locally to "establish clear expectations for hospitals using AI tools and better align payment with the accurate representation of care delivered." Expect the numbers to show up in state rate filings, employer premium conversations and federal AI policy debates. For AI vendors selling revenue-cycle and documentation software, the pitch of "capture every billable condition" is about to meet a customer on the other side of the claim that can measure it. The winning products will be the ones that can show the diagnosis and the treatment line up.