$942 million. That’s how much extra healthcare spending the Blue Cross Blue Shield Association says piled up over two years because hospitals started using AI tools when they submit insurance claims. Not new treatments. Not new machines. Not more patients getting more care. Just software getting better at describing what already happened.
If you’ve been reading about AI agents and wondering where they actually show up in ordinary life, here you go. Not a chatbot. Not a robot. A quiet piece of software sitting between a hospital’s billing department and an insurance company’s payment system, and a second piece of software on the other side arguing back.
What an AI agent is doing in a billing department
Medical billing runs on codes. Every diagnosis, every procedure, every complication gets translated into a standardized code, and those codes determine how much an insurer pays. A patient visit that gets coded as straightforward pays less than one coded as complex. That’s not fraud, that’s how the system was designed: sicker patients take more work, so more work should pay more.
Coding used to be done by people reading charts. It was slow, and hospitals routinely left money on the table because a human coder missed a detail buried in a doctor’s notes. AI tools read the entire chart, spot every documented condition, and suggest codes a person might have skipped.
From the hospital’s side, that’s just accuracy. They’re getting paid for care they actually delivered and previously under-billed. From the insurer’s side, the same patients are suddenly arriving with heavier, more expensive paperwork attached. The New York Times has reported on this as a growing problem, and the pattern it describes is the important part: billed complexity going up without a matching change in the care being delivered.
The part that makes this an arms race
Insurers are not sitting still. They have their own AI reviewing incoming claims, flagging the ones that look inflated, and generating denials. So hospital AI writes a more aggressive claim, insurer AI pushes back, hospital AI appeals. The New York Times framed it as hospital A.I. versus insurer A.I., and that framing is doing real work. This is a fight between two automated systems, each optimizing for its own side of a transaction.
For anyone trying to understand AI agents in plain terms, this is a useful example, because it shows something people often miss. An AI agent isn’t good or bad on its own. It’s a thing that pursues a goal very efficiently. Point one at “maximize reimbursement” and it will do that. Point another at “minimize payout” and it will do that too. Both are working correctly. Neither was ever asked to make healthcare cheaper or better.
When two optimizers grind against each other, the friction has to go somewhere. In this case it shows up as $942 million in extra spending, and spending in insurance eventually reaches you through premiums.
Why this pattern matters beyond healthcare
The healthcare version of this story is easy to follow because the money is measurable. But the underlying shape repeats anywhere two parties negotiate at scale.
- Any process where one side writes a request and the other side evaluates it
- Any process where the reward is winning the exchange rather than getting to the right answer
- Any process where both sides can automate faster than the rules can be updated
Job applications and resume screening already look like this. So do ad auctions. So do disputes and appeals of almost any kind. Once both sides deploy agents, volume explodes and the cost of participating goes up for everybody, including the people the system was built to serve.
A fair word for the hospitals
It would be easy to read the insurer numbers and conclude hospitals are gaming the system. Careful here, because insurers are the party publishing the analysis, and they have an obvious interest in the conclusion. If AI coding is genuinely catching conditions that human coders missed, then some share of that $942 million represents care that was always delivered and never paid for. There is no way to tell from the headline figure alone how much is correction and how much is inflation.
What we can say is that nobody chose this outcome deliberately. Two sides adopted useful tools for defensible reasons, and the combined result costs more than either intended.
What to take from it
When someone tells you an AI agent will make a process more efficient, ask a follow-up question: efficient for whom? Speeding up one side of a negotiation doesn’t make the negotiation better. It usually just makes the other side speed up too.
The interesting question for the next few years isn’t whether these tools work. They clearly do. It’s whether anyone is measuring what they cost the people standing outside the argument, holding the bill.
🕒 Published: