Dr. Bryant Lin teaches medicine at Stanford. He founded the Center for Asian Health Research and Education to study diseases that disproportionately affect Asian populations, including nonsmoker lung cancer. Then he was diagnosed with it himself. He is now Stage 4.
His doctor prescribed Rybrevant, an FDA-approved treatment for his specific mutation. Aetna denied the claim. Lin, a physician at one of the world’s most resourced medical institutions, had to beg for his life on LinkedIn. His post reached 500,000 people. Aetna reversed the denial.
Most patients don’t have half a million people watching. Most don’t appeal at all.
This is not a story about one insurer behaving badly. It’s a story about what happens when everyone in the system deploys AI to win a war, one that patients never signed up to fight.
Over 450 million claims are denied annually in the US. Payers have deployed proprietary algorithms, about which little is known, that never sleep, scanning for patterns, hunting discrepancies, automating denials at scale. Denial rates have climbed from 10% in 2020 to nearly 12% today—and higher for inpatient care.
The average US hospital loses $5m per year from rejected claims. So, unsurprisingly, on the provider side, hospitals have welcomed AI into revenue cycle management—for example, ambient documentation that transforms clinical encounters into optimized claims. Systems that study past denials to reverse-engineer payer logic, tools that adjust wording to avoid automated red flags. It’s driving a massive surge in adoption: 22% of US health providers have already begun to roll it out. When it comes to AI, an industry not known for its rapid implementation of operational change (confer the story of EHRs) is now in the lead.
The patient-side AI market is already emerging. Startups like Claimable now offer to craft appeal letters, citing policy violations, marshaling clinical evidence, even tuning the emotional register of the language. The reported success rates are impressive.*
But what does all this activity add up to?
Zoom out: we are now deploying AI systems to help patients fight AI systems that were built to deny them care, that other AI systems were built to bill for.
Capital is flowing not toward reducing friction, but toward intensifying it.
A recent report by Silicon Valley Bank calls this, explicitly, an “arms race.” They’re not wrong. But arms races have a particular economic logic: the returns go to the weapons manufacturers, not to the people caught in the crossfire.
The value of investment isn’t measured by whether you are “better off,” but whether you are less at risk than your rival. Much of the capital goes into duplication, not discovery. Everyone builds the same thing slightly differently, not because it’s needed, but because the rival has it.
From the point of view of vendors, none of this is a problem. They’re servicing customer demands—demands which their own customers agree to be urgent. But the overall effect of these rival deployments is yet to be seriously taken into account.
Provider AI is trained to extract value from the record. Payer AI is trained to protect patient value. Patients wait. Costs climb.
But none of this is fate.
Every metric is a belief system disguised as math. Every contract is a declaration of who must struggle and who must be spared. The negotiation has already begun; most people simply do not recognize that they are sitting at the table, that the terms can change.
We talk about AI adoption as if it were weather. Something that just happens, without opinion or intent. Every metric is a belief system disguised as math. Every contract is a declaration of who must struggle and who gets spared. Optimization is a choice – people are choosing what to optimize for. People are deciding the claims war matters more than spending that money on something that would result in care. Other agreements are possible. But first, we have to stop confusing inevitability with surrender.
This post is the first in a series. We’re going to map where else these adversarial dynamics are emerging—inside health systems, between clinicians and patients, across the boundaries of care. The arms race in the revenue cycle is just the most visible front.
This isn’t ultimately a story about AI. Or even about technology. It’s about embedded power dynamics within healthcare: who bears risk, who extracts value, who must prove their worthiness for care, are being encoded, accelerated, and legitimized through prevalent models of health systems. The ruptures we’re seeing aren’t bugs in the implementation. They’re revelations of the underlying architecture.