Health insurers say artificial intelligence is already driving up the cost of care in a way that has nothing to do with better treatment, and a new industry analysis is putting a number on it: $942 million in additional healthcare spending over a two-year period, attributed to AI-assisted documentation of insurance claims.

The figure comes from an analysis by the Blue Cross Blue Shield Association, which found a sharp increase in documentation of complex patient conditions coinciding with the spread of AI tools inside hospital billing departments. The association's analysis identified what it described as a disconnect: coding for more complex, higher-reimbursement conditions rising faster than any corresponding change in the actual medical treatment patients were receiving.

Two AI systems, working against each other

The dynamic driving the dispute is straightforward. Hospitals have increasingly adopted AI tools, including ambient documentation systems built by companies like Abridge, to help clinicians code patient visits more thoroughly and, in many cases, more accurately than manual documentation allowed. More complete documentation can capture legitimate clinical complexity that previously went uncoded, which in turn can mean higher, and arguably more accurate, reimbursement from insurers.

But insurers have simultaneously deployed their own AI systems to review claims, flag inconsistencies and push back on coding they judge to be inflated. The result, according to BCBSA senior vice president Luke Chalker, is an increasingly automated standoff between hospital-side coding AI and insurer-side claims-review AI, each system essentially optimizing against the other rather than against a shared, verifiable standard of what a visit actually involved. Chalker described the current dynamic bluntly: "It's not a war. It's a completely one-sided blood bath," arguing that AI-assisted coding is currently outpacing insurers' ability to verify it.

Dr. Shiv Rao, the physician who founded Abridge, one of the AI documentation companies at the center of the dispute, offered a starker warning about where the dynamic could lead if left unchecked, calling an escalating cycle of AI systems on both sides trying to outmaneuver each other "a horrible dystopic future nobody wants to live in" with "bots fighting bots."

Better documentation or upcoding, and who decides

The central disagreement is not really about whether AI-assisted documentation is happening. It clearly is, and at scale. The disagreement is about what to call it. Hospitals and AI documentation vendors argue that ambient AI tools are simply capturing clinical detail that overworked physicians previously left out of manual notes, complexity that was always present in the patient's condition but went undocumented and therefore unreimbursed. Insurers argue that at least some portion of the increase reflects AI systems finding technically defensible but clinically questionable ways to code visits at a higher complexity tier than the care actually delivered warrants, a pattern regulators have historically called upcoding when done by humans.

Both explanations are almost certainly true to different degrees across different hospitals and different AI tools, which is precisely what makes the dispute hard to resolve through a single audit or a single regulatory rule. Distinguishing legitimate AI-assisted documentation improvement from AI-assisted upcoding requires clinical chart review at a scale that, so far, only more AI seems capable of performing quickly enough to matter.

Why this is a governance problem, not just a billing dispute

For hospital executives and health system compliance officers, the BCBSA analysis is a warning that AI documentation tools, adopted primarily to reduce clinician burnout and improve note quality, are now generating a new category of financial and regulatory exposure. A hospital that cannot clearly explain why its AI-assisted coding shows a sustained increase in complex-condition documentation risks both payer pushback on individual claims and, potentially, broader scrutiny from regulators watching for systemic upcoding patterns.

For health-tech vendors, the dispute is a preview of a governance problem that will only get harder as AI tools take on more of the judgment calls, not just the transcription, involved in clinical documentation. An AI system that recommends a specific billing code based on a clinical note it also helped generate sits in a position that is difficult to audit cleanly, and difficult to defend if a payer or regulator asks the hospital to justify the recommendation independently of the tool that made it.

Sources