Picture an exam room. A doctor talks, a patient answers, and a microphone on the desk quietly turns the whole conversation into structured text. Somewhere in that transcript, software flags a detail the doctor might have summarized in one line and turns it into a documented, billable condition. The visit ends. The claim goes out. And on the other end of the wire, another piece of software reads that claim and starts looking for reasons to push back.
Nobody in that room chose this fight. But according to a report from the Blue Cross Blue Shield Association, it has already cost roughly $1 billion in extra spending over two years across insurers representing more than 100 million people.
What the report actually claims
The mechanism is less dramatic than the headline. Hospitals are using AI to scan medical records and transcribe patient encounters. Those tools surface more complex patient conditions than manual documentation typically captured. More documented complexity means higher severity coding, and higher severity coding means bigger reimbursements. The insurers’ position is that this is the tool working as designed, and the design happens to point money in one direction.
Insurers, meanwhile, run their own AI over incoming claims to scrutinize them. So you get two automated systems pointed at each other, each optimizing against the other’s output, with administrative cost piling up on both sides.
I’d note the obvious: this figure comes from insurers, describing a trend that costs insurers money. That’s not disqualifying, but it’s a number with a clear interest attached. Hospitals would frame the same activity as finally getting paid accurately for sick patients whose conditions were previously under-documented. Both can be partly true. Under-coding was a real problem before any of these tools existed.
Why this is a tooling story, not just a healthcare story
I review AI tools for a living, and this is the clearest real-world example I’ve seen of something I keep flagging in smaller contexts: a tool can perform exactly as promised and still produce a worse outcome for the system it sits inside.
The ambient documentation tools in question aren’t broken. By the report’s own account, they find things humans missed. That’s the feature. Every vendor demo in this category sells the same pitch — less typing, better notes, more complete records. On the metrics the buyer cares about, the tools deliver.
The buyer’s metrics just aren’t the system’s metrics. And when an AI tool is deployed into an adversarial process, “better at the task” collapses into “better at winning.” Documentation isn’t a neutral record when it determines payment. It’s a negotiating position. Point a capable model at a negotiating position and it will strengthen it.
The part that should worry you
The escalation loop is the real finding here. Hospital software gets better at documenting severity. Insurer software gets better at questioning documentation. Each side’s improvement justifies the other side’s next upgrade. Neither side can stop unilaterally, because stopping means losing money to a counterparty that didn’t stop.
This is what an automation arms race looks like when nobody is measuring total system cost. The spending goes up on claims. It also goes up on the tools themselves, the integration work, the appeals staff, the audit teams. None of that shows up in the vendor’s ROI calculator, which counts minutes saved per clinician and stops there.
And the person paying for both sides of this fight, eventually, is the one in the exam room who never saw either tool.
What I’d take from this if you’re buying AI tools
The lesson generalizes well beyond healthcare, and it’s worth sitting with before your next procurement cycle:
- Ask who the tool is optimizing against. If the answer is another party’s software, expect an arms race rather than a one-time efficiency gain.
- Distrust ROI framed only as time saved. Minutes reclaimed is the easiest number to produce and the least useful one. Ask what downstream costs the tool creates for anyone else.
- Watch for tools that change outputs while claiming to only change process. “We just transcribe the conversation” is not a neutral act if the transcript determines a payment.
- Budget for the counter-tool. In any adversarial workflow, adopting AI on one side reliably summons AI on the other. That’s a cost line, not a surprise.
The uncomfortable read on all this is that both sets of tools are doing good work. The documentation AI is finding genuine complexity. The claims-review AI is catching genuine overreach. Put them in the same room and the combined result is a billion dollars of friction and no obvious improvement in anyone’s health.
That’s the thing about deploying capable software into a system with misaligned incentives. It doesn’t fix the incentives. It just makes them faster.
đź•’ Published: