AI Coding Tools: The $942M Insurer Claim

A patient comes in with a hip fracture. The chart says hip fracture. Then the software reads the labs, the notes, the nursing entries, and suggests four more diagnoses. Anemia. Electrolyte disturbance. Something with a longer name. A clinician clicks accept. The claim moves up a tier.

Nobody did anything dishonest in that scene. That is the problem.

What the insurers say

On September 26, the Blue Cross Blue Shield Association published an analysis of its own claims data. Its headline figure: hospitals' use of AI coding tools added an estimated $942 million in costs (as reported by TechCrunch). The period was 2023 to 2025, per Health System CIO.

The mechanism, per the same coverage: more patients are documented with complex conditions, and BCBSA found "no evidence of corresponding change in care delivered." About 70% of the added cost, roughly $650 million, came from secondary diagnoses.

This is an insurer's analysis. TechCrunch notes the claims were presented without independent verification of the methodology or a detailed hospital response. Read it as an accusation, not a verdict. We could not load the primary BCBSA page (it returned an error), so we lean on the two secondary reports above.

What the hospitals say

The hospital side is blunt. Julie Demaree of St. Mary's Healthcare said insurers "have used technology to review and deny claims for years," and that AI documentation tools "help clinicians record how sick a patient is." James Wellman of Nathan Littauer Hospital called adoption necessary for survival: "Without it, you're just going to hold off the inevitable."

Both arguments can be partly true. Under-documentation is real. Sicker patients have long been recorded as less sick than they are. A tool that catches a missed diagnosis is doing something useful. A tool tuned to find every billable code is doing something else.

Why this lands on the ward

Doctors are the ones who accept the suggestions. The chart carries the clinician's name. If a payer audits it, the question comes to whoever signed.

Ochsner Health's Jason Hill made the point that a good audit trail records how much of a note was generated by AI and how much came from the ambient conversation. Health System CIO adds the caveat that matters: clinician approval alone does not establish clinical validity. A click is not a clinical judgment.

That is the gap. Ambient scribes and coding assistants sit between the patient and the record. When they work well, the record is closer to the truth. When the incentive is revenue, the record can drift from it, one accepted suggestion at a time.

What we take from it

We build for the individual doctor on rounds, so our view is narrow. Three things.

1. Provenance should be visible. Every line in a note should show where it came from: the doctor, the conversation, the lab, or a model. If you cannot tell, you cannot defend it.
2. Suggestions should point at evidence. A proposed diagnosis with the value and the timestamp behind it is checkable in seconds. One without is a request for trust.
3. The doctor's tool should serve the doctor's patient. A system built to maximise the claim and a system built to support the round are different products, even when they read the same chart.

None of this settles the $942 million question. That will be argued between payers and hospitals for a while. But the signal for clinicians is clear: what you accept in the chart is now a financial and legal object, not only a clinical one.

Sources

• Insurers claim AI is already increasing healthcare costs, TechCrunch, Sept 26, 2026
• Insurers Blame Hospital AI Coding for $942 Million in Added Costs, Health System CIO, Sept 29, 2026
• BCBSA analysis page (referenced by TechCrunch; not independently loaded)