In early 2026, the CEOs of nearly every major health insurer sat on earnings calls and told Wall Street analysts roughly the same thing: we’re going to cut costs, and AI is part of how we do it.
I build bots for a living, so my first reaction wasn’t about healthcare policy. It was recognition. That sentence is what every engineer says right before they deploy an automated system into an environment where somebody else also has an automated system. And the reporting coming out now says the predictable thing happened. Hospitals pointed AI at documentation and coding. Insurers pointed AI at prior authorization and claims review. The bills went up.
Payers are already finding that AI-driven documentation and coding increases billing amounts. Hospitals using AI to capture more detailed patient information also pushes costs higher. The New York Times framed it as a battle of hospital AI versus insurer AI, escalating a feud that predates both systems by decades. The AI didn’t create the conflict. It just gave both sides a much faster trigger finger.
This is an adversarial loop, not an efficiency gain
Here’s what I keep thinking about as someone who writes automation for a living. Most of us build bots with an implicit assumption baked in: the environment is passive. You scrape a page, the page doesn’t scrape back. You classify a ticket, the ticket doesn’t reclassify itself in response.
Healthcare billing breaks that assumption completely. There are two parties with directly opposed objective functions operating on the same documents. One side optimizes for maximum defensible reimbursement. The other optimizes for minimum defensible payout. Drop capable automation on both ends and you don’t get a settled equilibrium, you get an arms race with a much shorter cycle time.
Anyone who has built trading bots, ad bidding systems, or anti-fraud pipelines has watched this movie. The pattern is consistent:
- Side A automates and gains an edge, because side B is still human-paced.
- Side B automates in response, and the edge evaporates.
- Both sides now run more volume, more complexity, and more review cycles than before.
- Total system cost goes up even though each individual tool is doing exactly what it was built to do.
Every component passes its unit tests. The system as a whole gets worse. That’s the part that should bother us professionally, not just as patients.
Documentation tools are optimization engines wearing a helper costume
The framing around clinical documentation AI has mostly been about relieving burden. Less typing, fewer late nights finishing notes. That’s a real problem worth solving, and I’d happily build for it.
But a tool that suggests more complete coding is, mechanically, a tool that increases revenue capture. Those are the same function described two different ways. Analysts tracking 2026 cost trends put AI documentation and coding tools on the list of factors likely to accelerate rising costs, alongside an increasingly complex patient population. When your assistive feature and your revenue feature are the same code path, the product roadmap only ever points one direction.
Insurers are doing the mirror image with prior authorization, which has become one of the main places AI shows up in payer workflows. More coverage decisions run through automated systems, and the risks to patients grow when the thing deciding your care is a model optimizing a cost target.
The uncomfortable question for builders
If you ship a system whose measurable success looks identical to a worse outcome for the people it touches, what exactly did you validate? I don’t have a clean answer, and I’m suspicious of anyone who says they do. But I’ve started asking a few things before I write the first line of a bot that operates in a contested space:
- Is there an opposing party with an opposed objective? If yes, assume they automate too, and model the second-order state.
- Does my metric measure the outcome, or measure my side winning? Throughput and approval rates are side-winning metrics.
- What happens to total system cost if my tool succeeds completely? If that number goes up, “efficiency” is the wrong word for what I built.
- Who absorbs the cost of the loop? In healthcare, it isn’t the hospital or the insurer. It’s the person holding the premium notice.
Where this leaves us
AI’s role in escalating medical expenses is now a genuine point of contention, with each side blaming the other’s software. Both are probably right, which is the whole problem.
The lesson generalizes past healthcare. Fast automation in an adversarial setting doesn’t reduce friction, it industrializes it. If you’re building agents that negotiate, adjudicate, price, or contest anything against another party’s agents, you’re not building a productivity tool. You’re building one half of a feedback loop, and the loop is the real product. Design accordingly, and be honest in the design doc about who pays for the escalation.
🕒 Published: