The first real-world, at-scale AI productivity win is here, and it’s an arms race over who gets to keep your money.
Blue Cross Blue Shield says hospital use of AI tools added roughly $942 million in costs over two years. Hospitals are running AI to document and code encounters, those encounters come back billed higher, and insurers are running their own AI to push back. That’s the whole story. Two automated systems arguing over an invoice, with the patient standing outside the room.
I review AI tools for a living. I read a lot of vendor decks promising efficiency. This is what efficiency looks like when nobody defines what it’s efficient at.
What the tools actually do
Clinical documentation and coding software is one of the few genuinely good fits for language models. Messy notes in, structured codes out. It’s repetitive, it’s rules-heavy, and humans hate doing it. On paper, this is the boring back-office automation everyone said would arrive before the sci-fi stuff.
The catch is what the tool is optimized for. A coding assistant isn’t graded on accuracy in any abstract sense. It’s graded on capture — finding the billable detail a tired human would have skipped. Every vendor in that category sells on revenue lift. That’s the pitch, right there in the marketing. Nobody is selling hospitals a tool that finds less money.
So when payers report that AI-driven documentation and coding is raising billing amounts, that isn’t a malfunction. That’s the product working. The software did the job it was bought to do.
Neither side is the good guy
It would be easy to read the insurer complaint as a story about hospitals gaming the system. Resist that. Insurers are deploying their own AI on the other end of the pipe, automating review and denial at a speed no human reviewer could match. The feud between hospitals and insurers over who pays for what is decades old. AI didn’t create it. AI just gave both sides a much faster trigger finger.
And the costs of that escalation don’t vanish. Every automated claim that gets automatically denied and automatically resubmitted burns real money on both ends. Administrative overhead is already one of the most expensive absurdities in American healthcare. We just bought it a turbocharger.
The $942 million figure comes from one side of a fight, so treat it as a claim rather than a settled number. Insurers have an obvious interest in framing hospital AI as the cost driver. But the underlying mechanism is not in dispute, and the direction of travel is not surprising to anyone who has looked at how these tools are sold.
What this means for the rest of us
I spend most of my time telling readers which AI tools are worth the subscription. The lesson here scales past healthcare, and it’s one I keep running into with agents in general.
- An AI tool inherits the incentives of whoever bought it. Ask what metric the vendor optimizes for, because that’s the behavior you’re buying. Coding tools that promise revenue capture will capture revenue.
- Automation makes adversarial processes worse, not shorter. When two parties with opposed interests both automate, you don’t get resolution. You get volume.
- Efficiency gains can land entirely outside the system they were measured in. The hospital’s billing department genuinely got more efficient. Total cost went up anyway. Both things are true.
- “The AI did it” is becoming an accountability laundering service. A human upcoder is a fraud investigation. A model that produces the same output is a software configuration.
That last point is the one I’d watch. The tools are new enough that nobody has settled who owns the output. If a coding model consistently produces higher-value claims, is that the hospital’s decision, the vendor’s default setting, or an emergent property of the training data? Everyone involved has a reason to prefer the vaguest possible answer.
My honest read
This is the clearest example yet of AI delivering exactly what was promised and making things worse. No hallucination scandal, no rogue agent, no safety incident. Just software that does its narrow job well inside a system where doing that job well is the problem.
If you’re evaluating AI for your own operation, the useful question isn’t whether the model is accurate. It’s whether the thing you’re automating should go faster. Healthcare billing disputes were already too expensive, too slow, and too hostile. Adding models to both sides didn’t fix any of that. It scaled it.
Hospitals will keep buying capture tools because they work. Insurers will keep buying denial tools because those work too. Both vendors get paid either way. Patients get the bill, and the industry gets to describe the whole arrangement as modernization.
Good tools. Bad system. The tools won.
🕒 Published: