Remember when the big promise of AI in healthcare was that it would finally get doctors out from under the mountain of paperwork? The pitch was lovely. Let the software write the notes, and physicians get their evenings back. Nobody in that pitch meeting mentioned your premium going up.
That’s the twist insurers are now pointing at. Blue Cross Blue Shield has said its data backs up the claim that AI is driving up medical bills, according to reporting from STAT. The mechanism is not some sci-fi scenario. It’s documentation. AI tools help providers write better, more detailed notes, and more detailed notes support higher complexity codes, and higher complexity codes mean bigger bills. Payers say they are already seeing billing amounts climb as a result, and bigger bills eventually show up as higher premiums.
How a note becomes a number
If you’ve never had to think about medical coding, here’s the short version. After a visit, someone translates what happened into standardized codes. Those codes carry different payment weights depending on how complex the encounter was. A quick, simple visit codes lower. A visit involving multiple conditions, careful decision-making, and a lot of documented reasoning codes higher.
Now add software that is genuinely good at capturing detail. An AI scribe listening to a visit doesn’t get tired, doesn’t skip the third chronic condition the patient mentioned in passing, and doesn’t leave out the reasoning behind a treatment decision. All of that detail is real. It also happens to be exactly the kind of detail that supports a higher-complexity code.
So the interesting question is whether this is upcoding or accurate coding. Insurers use the word upcoding, which implies bills inflated past what the care justified. Providers can reasonably answer that they were undercoding for years because humans are bad at documentation and nobody has time to write everything down. Both things can be partly true at once, and that ambiguity is the whole fight.
What this means for AI agents specifically
This is one of the clearest real-world examples of something we talk about a lot on this site. An AI agent is software that takes actions toward a goal on someone’s behalf. Give an agent a goal, and it will optimize for that goal with a thoroughness no human employee sustains across ten thousand cases.
A documentation agent pointed at “capture everything clinically relevant” will do that relentlessly. It is not scheming. It is doing the job. The cost effect is a side effect of the job being done well, and that’s the part organizations keep underestimating when they deploy these tools:
- The agent’s goal was documentation quality, not spending control.
- Nobody wrote “and don’t move the billing average” into the instructions.
- The effect only becomes visible in aggregate, across millions of claims.
- By the time it’s visible, it’s a trend, not a bug you can patch.
That last point matters. A single note coded one level higher is invisible. The same shift across a health system, month after month, is a line on a financial report.
And then the other side gets agents too
Insurers are not passive here. A KLRD briefing on AI use in health insurance for 2026 notes that one of the main ways insurers have adopted AI is in prior authorization, the process where providers have to get approval before certain care goes ahead. CBS News covered the broader picture in March 2026, framing it as hospitals and insurers both turning to AI in fights over claims and payments.
So picture the setup. Provider-side software builds the strongest possible documented case for a claim. Payer-side software evaluates, questions, and approves or denies at machine speed. Two sets of agents, opposing goals, processing volumes no human team could touch.
The word for that is not efficiency. It’s an arms race, and arms races consume resources rather than freeing them. Administrative cost in American healthcare was already famous for being enormous before either side had automation. Making both sides faster doesn’t obviously shrink it.
Where patients sit
Awkwardly, is the honest answer. Insurers say they’re adjusting policies to manage the spending they’re seeing, and policy adjustments are the thing patients actually feel. Tighter rules, more approval steps, different coverage terms. A report on trends shaping 2026 healthcare costs lists AI documentation and coding tools as likely to accelerate cost pressure alongside factors like an increasingly complex patient population.
None of this makes AI documentation tools bad. Better records can mean better care and fewer things falling through cracks. But it’s a useful correction to the assumption that automation automatically means savings. Automation means whatever the goal you gave it means, executed at scale.
The lesson generalizes well beyond medicine. Before you deploy an agent, be honest about what you actually told it to optimize for, because that is what you will get. Lots of it. Faster than you expected.
🕒 Published: