Essays · AI & the economics of getting paid

The AI Arms Race Nobody Voted For

By Zack Spooner, founder of Remend Health and an AI engineer by trade. Third in a short series on AI and the economics of getting paid in American healthcare.

For the last several years, the automation in American claims processing ran in one direction. Payers built systems to deny, adjust, and downcode at machine scale. Providers responded with people: billers, coders, and appeals staff working claims by hand. I have written in this series about what that asymmetry produced, denials issued in 1.2 seconds and underpayments calibrated to sit just below the threshold of being worth a fight.

That one-sidedness is ending. And the thing replacing it is not obviously better.

Both sides are now automating

In January 2026, Health Affairs published a peer-reviewed analysis by a team at Stanford with a title that says the quiet part out loud: "The AI Arms Race in Health Insurance Utilization Review." The framing is no longer speculative. Insurers and provider organizations are both escalating their use of AI in prior authorization and claims, each responding to the other, and the paper's warning is specific. When both sides automate a process that was already flawed, you do not get a fair fight. You get what the authors call "supercharged flaws," the same errors as before, now executed faster and at greater scale, with humans pushed further from the loop on both ends.

You can see it happening in the market. A former chief data officer at Kaiser, UnitedHealthcare, and Optum told Healthcare Brew in October 2025 that insurers are now the ones feeling pressure, because providers adopted AI faster than expected and left payers with, in his words, old technology and humans responding to deluges. Survey data backs the shift. Bain and KLAS found that 70 percent of providers now have an AI strategy in place or in development, with revenue cycle work ranking as a top-three IT priority for nearly half of them. A separate HFMA survey put the share of health systems exploring, piloting, or deploying generative AI in revenue cycle at 80 percent.

So the appeals letter that used to take a biller an hour now drafts itself. The claim that used to be written off as too small to chase now gets flagged automatically. The denial that used to stand because no one had time to fight it now gets contested by a system that never runs out of time. On its face, that sounds like the asymmetry I have been describing finally correcting itself.

I am not sure it is that simple.

What an arms race actually produces

The optimistic read is that provider-side AI rebalances the scales. Denials get appealed, downcoding gets caught, and payers, facing a counterparty that can now respond at their own scale, deny less aggressively because the easy denials no longer stick. That would be a genuinely good outcome, and parts of it may happen.

But arms races do not usually end with one side winning. They end with both sides spending more to hold roughly the same ground. There is a real risk that what we are building is a closed loop where one company's model generates a denial, another company's model generates an appeal, the first model adjudicates the appeal, and the volume of automated traffic between them climbs every quarter while the underlying question, did this patient get the care they needed and did the doctor get paid for it, gets answered by fewer and fewer humans.

The Stanford authors are pointed about the danger here. Automation bias, where reviewers defer to whatever the model outputs, can hollow out the human review that is supposed to be the safeguard. Opacity on both sides makes it harder, not easier, to tell whether any individual decision was correct. An arms race optimizes for throughput. It does not optimize for being right.

Where I land

I think the automation of the provider side is inevitable and, on net, necessary. The status quo, where payers operate at machine scale and providers at human scale, is not a fair equilibrium and it is not one worth defending. If a practice is going to be downcoded by an algorithm, it should not have to answer with a tired biller and a fax machine. Leveling that up is the right thing to build.

But "necessary" is not the same as "sufficient," and "faster" is not the same as "fair." The version of this that is actually good for patients is not just more AI on the provider side. It is provider-side AI that makes correct claims harder to wrongly deny, that surfaces the underpayments that should never have happened, and that does it transparently enough to withstand the scrutiny the payer systems have mostly avoided. The version that is bad for everyone is two black boxes trading automated paperwork while the patient waits.

We did not vote for this arms race. It is happening anyway. The only real choice left is what kind of automation the provider side brings to it, and whether it is built to win the volume war or to actually get the answer right. Those are not the same goal, and I think the next few years come down to which one the builders choose.