The Blue Cross Blue Shield Association - 31 independent insurers covering more than 100 million people - published a study on 24 September 2026 finding that AI-assisted clinical documentation has added roughly $942 million in costs measured against a 2023 baseline.
This is the first large-scale look at what happens when ambient AI scribes meet a reimbursement system that pays more for complexity, and the mechanism it describes is worth understanding whether or not you work in healthcare.
The Numbers
- $942 million total attributed to coding and classification changes versus 2023
- $653 million of that from secondary diagnoses between 2024 and 2025
- Study window: 2023 to 2025
- Coverage base: 31 Blue Cross companies, 100+ million members
The Mechanism
Ambient AI scribes sit in the consultation room, transcribe the doctor-patient conversation and draft the clinical note automatically. Some also scan the medical record for conditions that were not explicitly discussed.
They are genuinely useful. Clinical documentation is the most-cited cause of physician burnout, and a tool that removes two hours of evening typing is not a scam.
But US reimbursement pays by documented complexity. A visit recorded as more complex qualifies for a higher payment. So a tool that documents more thoroughly and a tool that bills more are, structurally, the same tool. Nobody has to intend anything for the bill to go up.
The Finding That Makes It a Story
Among patients having major bowel surgery, between Q1 2023 and Q4 2025:
- Partial intestinal blockages: up 55%
- Excess acid diagnoses: up 33%
- Anemia diagnoses: up - blood transfusions: unchanged
That last line is the whole argument. If patients were genuinely sicker, treatment volumes would move with the diagnoses. Diagnoses rose. Treatment did not.
Luke Chalker, senior vice president at BCBSA, put it plainly: "The disconnect between diagnoses and treatment suggests that AI is identifying more billable conditions, not sicker patients."
Centene has raised similar concerns separately.
Read the Study Carefully Too
This is an insurer study about costs insurers pay, and that is a position, not a neutral vantage point. A few honest caveats:
- Under-coding was real. Some of this increase is genuine conditions that were previously missed because a tired clinician did not write them down. Better documentation finding real comorbidity is a good outcome, and the study cannot fully separate that from the rest.
- The treatment gap is suggestive, not conclusive. Not every documented condition warrants a treatment. Anemia can be noted and monitored.
- No sample size was published alongside the headline figure, which is the first thing to ask for.
The diagnosis-without-treatment pattern is still the strongest evidence in the study, and it is hard to explain away at 55%.
The Generalisable Lesson
Strip out the healthcare specifics and this is a story about an AI tool optimising exactly what it was pointed at, in a system where the measured thing and the valuable thing had quietly come apart.
Nobody asked the scribe to increase billing. It was asked to document thoroughly. Thorough documentation happens to be what triggers higher payment, so the bill went up, and it will keep going up until someone changes either the tool or the incentive.
That shape shows up wherever AI meets a metric with money attached to it. If you are deploying AI against any process where a number determines a payment - claims, procurement, performance, commission - this study is the cautionary version of your own roadmap.
What To Do About It
- Measure the downstream, not the output. BCBSA found this by checking whether treatment followed diagnosis. The equivalent question for your deployment is whether the thing the AI produces more of leads to the thing it was supposed to cause.
- Baseline before you deploy. The only reason a $942 million figure exists is a 2023 comparison point.
- Assume the tool optimises the measured thing. Not the intended thing. Where those differ, the gap is your bill.
Sources