$942 million. That’s the extra cost Blue Cross Blue Shield attributes to hospitals using AI tools, tallied over two years. Not a forecast. Not a projection for 2030. Money already spent, according to the insurer, because software got better at writing up what happened in an exam room.
If you’ve been waiting for a real-world example of AI agents doing something consequential outside a demo video, this is it. It’s just not the example anyone was hoping for.
What’s actually happening here
Let me explain the mechanics, because they’re simpler than they sound.
When you visit a hospital, someone has to translate your visit into codes. Codes describe your diagnosis, the complexity of your case, the procedures performed. Those codes determine what the hospital gets paid. A visit coded as straightforward pays less than one coded as complicated.
Historically this translation work was done by humans reading doctors’ notes. Human coders are inconsistent, get tired, and sometimes miss details that would have justified a higher payment. Hospitals lost money to that sloppiness for decades.
Now AI documentation and coding tools read the notes instead. They don’t get tired. They don’t miss things. They catch every detail that supports a higher-paying code, on every single claim, forever. Payers are finding that this AI-driven documentation and coding is increasing billing amounts.
The hospitals aren’t necessarily doing anything wrong. That’s the uncomfortable part. Their software is just very, very good at its job.
Then the insurers built their own
Insurers responded the way you’d expect. They deployed AI to review claims, spot patterns, and push back on the ones that look inflated.
So now there are agents on both sides of the table. Hospital software writes the claim to maximize payment. Insurer software reads the claim to minimize it. The result, as the reporting describes it, is a battle between hospital AI and insurer AI that’s pushing medical costs higher rather than lower.
The feud between hospitals and insurers over who pays what is decades old. AI didn’t create it. AI just handed both sides a weapon that never gets tired, never takes a lunch break, and gets cheaper every quarter.
Why this matters for how you think about AI agents
On this site I spend a lot of time explaining what AI agents are: software that takes goals, makes decisions, and acts without a human approving every step. Most examples are friendly. An agent that books your travel. An agent that sorts your inbox.
The healthcare billing fight is a better teacher than any of those, for three reasons.
First, agents optimize for what you tell them, not what you meant. A hospital tells its coding agent to capture every justifiable dollar. The agent does exactly that. Nobody instructed it to consider whether total system costs go up. That wasn’t in the goal, so it isn’t in the behavior.
Second, agents on opposite sides don’t cancel out. You might assume hospital AI and insurer AI would balance each other into a fair number. Instead both sides now spend money on software, staff to run it, and the disputes it generates. The friction is real work, and real work costs money. Somebody pays for that, and it isn’t the software.
Third, speed changes the nature of a conflict. A disagreement that used to move at the pace of paperwork now moves at the pace of servers. Arguments that would have been dropped because they weren’t worth a human’s afternoon now get pursued automatically, at volume.
What I’d watch next
I’m not going to pretend I know where this ends. But the pattern is worth recognizing, because it will repeat anywhere two parties already disagree about money and both get access to capable automation. Loan applications. Insurance claims of every kind. Contract disputes. Tax filings.
The useful question isn’t whether AI makes any single organization more efficient. Both hospitals and insurers are almost certainly more efficient at their specific jobs than they were three years ago. The question is whether the system they share got cheaper. In this case, the insurer’s own number says no.
That distinction matters when someone pitches you an AI agent. Efficient for whom, measured against what, and who absorbs the cost of everyone else getting efficient at the same time?
The part that should bother you
Nobody in this story is a villain. Hospitals want accurate payment for real care. Insurers want to stop paying for inflated claims. Both are using software that does what software is supposed to do.
And costs went up by $942 million anyway.
That’s the thing about agents worth understanding before you deploy one: they’re extremely good at winning the specific game you point them at. Whether that game was worth winning is a judgment call, and it’s still yours to make.
🕒 Published: