Two facts that do not want to sit in the same room together. Fact one: AI documentation and coding tools do what they promise. They catch detail that human notes miss, they translate messy clinical reality into accurate codes, and providers who adopt them report better records. Fact two: insurers now say those same tools are driving up medical bills and, by extension, your premiums. Blue Cross Blue Shield has gone on record saying its data backs the claim that AI is pushing medical costs higher.
Nobody is arguing the tools are broken. That is what makes this interesting, and what makes it worth talking about on a site that reviews AI tooling for a living.
What the insurers are actually alleging
The claim is upcoding. AI-assisted documentation produces richer notes, richer notes support higher complexity codes, and higher complexity codes mean bigger bills. Payers say they are already seeing billing amounts climb as these tools roll out. Industry forecasting on 2026 healthcare costs lists AI documentation and coding among the trends expected to accelerate spending, alongside a more complex patient population. Insurers have started adjusting policies in response, which is corporate language for tightening the valve on what gets paid.
Note the word “irrespective” hanging in that forecast language about billing increases. The payer position is not simply that AI helps doctors cheat. It is that the output goes up whether or not the underlying care changed. That distinction is the entire fight.
Both sides can be telling the truth
I test tools for a living, and this pattern is familiar. A tool that surfaces information more completely will change the numbers downstream. If clinicians were historically under-documenting, then better documentation produces higher codes that were always justified. The bills go up. Nothing improper happened. The previous number was simply wrong in the other direction.
But the same mechanism supports a less flattering reading. A tool optimized to find every billable element will find every billable element, including ones a cautious human coder would have left alone. Vendors sell revenue lift as a feature. When the pitch deck promises higher reimbursement per visit, you cannot be shocked when payers treat the results as a billing strategy rather than a documentation improvement.
From the outside, these two scenarios produce identical data. Codes go up. That is why this is turning into a data war instead of a conversation.
The arms race nobody planned
CBS News has reported on hospitals and insurers both turning to AI in disputes over claims and payments. Providers use AI to document and code. Payers use AI in prior authorization to evaluate and deny. Legislative briefing material on AI in health insurance points to prior authorization as one of the primary places insurers have deployed it.
So we now have automated systems generating claims and automated systems reviewing claims, escalating against each other at machine speed, with a patient somewhere in the middle waiting to find out whether their procedure is covered. Neither side built these tools to serve that patient. Both built them to win a financial argument. The patient absorbs the outcome through premiums, denials, and paperwork.
What this means if you are evaluating these tools
A few things I would want to know before adopting anything in this category:
- How does the vendor describe success? If the primary metric is revenue per encounter rather than accuracy or clinician time saved, you are buying a billing optimizer. Be honest with yourself about that.
- Can you audit the code suggestions? A tool that explains why it landed on a complexity level gives you something to defend. A tool that just outputs a code leaves you exposed when a payer pushes back.
- Does your coding distribution shift after deployment? Track it. If your case mix did not change but your code mix moved sharply, you want to understand that before an insurer does.
- What happens when payer policy changes? Insurers are already adjusting. A tool tuned to today’s reimbursement rules is a depreciating asset.
The uncomfortable part
The broader lesson here goes past healthcare. We keep evaluating AI tools on whether they perform the task well, and we keep getting surprised when tools that perform the task well produce system-level effects nobody wanted. A documentation assistant that accurately captures clinical complexity is a good tool. Thousands of them running across a payment system built on the assumption of imperfect documentation is a different thing entirely.
Insurers are not neutral narrators here. They have obvious incentive to blame rising costs on provider behavior, and prior authorization automation deserves at least as much scrutiny as coding automation. But the mechanism they are describing is real and easy to understand, and the people selling these tools have been advertising exactly that mechanism as a selling point.
The tools work. That was never the question. The question is what happens when everyone has them, and so far the answer looks like a more expensive standoff with the same patients paying for it.
🕒 Published: