AI Medical Billing Added $942 Million. The Care Stayed the Same.

Hospitals' AI note-taking tools are linked to $942 million in additional healthcare spending over 2 years, and the care patients received did not change to match. That is the finding of the Blue Cross Blue Shield Association, which represents 31 independent Blue plans covering more than 100 million Americans. Its analysis describes "a sharp increase in patients being documented as having complex conditions" with "no evidence of corresponding change in care delivered."

The tools at the center of this are ambient scribes: software that listens to a patient visit and drafts the clinical note on its own. The reporting names no hospital and accuses none of fraud. Hospitals adopted these tools for a sound reason, since doctors were spending their evenings on paperwork and a scribe hands some of that time back. What BCBSA found is AI medical billing colliding with oversight built for the pace of human documentation. The $942 million is the price of that collision.

The record says sicker, the treatment says the same

Of the $942 million, $653 million (about 7 of every 10 dollars) ties to hospitals billing more frequently for secondary conditions. These are the extra diagnoses recorded beside the main reason for a hospital stay, and they raise what a claim is worth. Major bowel surgery shows the pattern most clearly. Between the first quarter of 2023 and the last quarter of 2025, documentation of secondary conditions such as partial intestinal blockages and excessive stomach acid rose substantially for those procedures.

Luke Chalker, BCBSA's senior vice president of product and data science, put the whole puzzle in one line:

"If patients are truly sicker, we'd expect to see more treatment."

Consider it from the patient's side of the bed. Every one of those added diagnoses belongs to a real person who recovered from surgery and went home. Many of them will never read the note that describes them. The note travels anyway, into the next claim and into the premiums their employer negotiates next year.

Audits built for a human pace

The Centers for Medicare & Medicaid Services writes much of the coding rulebook that the rest of American healthcare builds on, commercial insurers like the Blue plans included. Its National Correct Coding Initiative has run since 1996. It was an early, serious piece of automation: a published set of edits that catches code combinations that should not appear together on a claim. Around it sit sampled claims reviews designed to catch coding drift. Drift is the slow creep that builds when thousands of human coders make slightly different judgment calls over the years.

That design fits the world it was built for. A human coder working from a physician's note produces documentation at a human pace, and drift appears slowly enough for sampling to find it. An ambient scribe documents every visit with the same exhaustive attention, around the clock. When it learns to record a secondary condition, it records that condition across a whole hospital system at once. These are rules written for a slower world, and the pattern BCBSA describes took shape in under 3 years.

Even the people building these tools can see where an unmanaged version leads. Dr. Shiv Rao, founder of the clinical-documentation company Abridge, has acknowledged the risk of "a horrible dystopic future" in which competing AI systems game each other, while arguing that AI could lower costs over the long run. I take both halves of that seriously. If hospitals document with AI and insurers answer by auditing with AI, the result is 2 machines arguing over a patient neither one is treating.

The learning debt comes due

In The Great Relearning, the book I am writing now, I call this kind of cost the learning debt: the modernization an institution defers because the old system still works well enough, until the bill arrives at once. Here the debt belongs to the whole system: hospitals, insurers, software vendors, and the oversight built around all of them. That oversight was calibrated for human documentation because human documentation was all there was. The $942 million is the first statement.

The same pattern appears inside companies. When I wrote about why companies break the AI rules they wrote, an EY survey showed 98% of large US companies with formal AI governance policies and 47% bypassing them under deadline pressure. And New York City's proposed AI kill switch and whistleblower bounty is a city trying to set its rules before the technology sets the pace. More on this pattern lives in my AI governance writing.

Where the next move belongs

This is exactly the kind of pattern a coding-integrity rulebook built for AI-scale documentation could get ahead of, and CMS is well placed to write it. The work would look familiar to anyone who knows the National Correct Coding Initiative: automated, published checks, this time comparing what the note says with what the care shows. When documented complexity rises and treatment stays flat, the check flags it, which is the same test Chalker applied. Add a simple disclosure on claims drafted with AI assistance, and the audit system starts reading at the speed the documentation is written.

I advise the California State University system on AI governance, and the question we keep returning to there applies here too: when a tool changes how fast work happens, which checks have to change with it? The answer is rarely to slow the tool down. Doctors deserve their evenings back, and patients deserve records that describe them accurately. Both can be true once the checks move at the same speed as the notes.

If you lead a hospital or a health plan, take one question into your next leadership meeting: does the complexity our AI documents match the care our people delivered? And if you are a patient, ask for your visit notes and read them. The record is yours.

For decades the clinical note was written at the speed of a tired doctor at 9 p.m. Software writes it now. The rules can move to that speed too, and the people who write them will find plenty of hands willing to help. Mine is one of them, and I'm easy to find.

Here is what makes Alex a credible voice on this topic: Alex shaped Cisco's $1.1 billion innovation portfolio for nearly two decades. That work taught him oversight has to move as fast as the technology it governs, and he now advises the California State University system on AI governance for 460,000 students.

Want help building oversight that keeps pace with your AI rollout? Bring Alex in to help →