Essays · AI & the economics of getting paid

The 1.2-Second Denial

By Zack Spooner, founder of Remend Health and an AI engineer by trade. I write about where AI is actually changing healthcare, not where the press releases say it is.

In 2022, over a single two-month stretch, Cigna's medical directors rejected more than 300,000 requests for payment. The average time a doctor spent reviewing each one was 1.2 seconds.

That number comes from a ProPublica and Capitol Forum investigation, and it is worth sitting with. 1.2 seconds is not enough time to open a patient's file, let alone read it. A former Cigna doctor described the workflow plainly: "We literally click and submit. It takes all of 10 seconds to do 50 at a time." The system, internally called PxDx, batched claims by diagnosis code and let physicians sign off on denials in bulk without ever seeing the underlying records.

I spent the last stretch of my career building AI systems in healthcare, and I have come to believe the denial machine is the most under-discussed AI story in the industry. Everyone is talking about ambient scribes and chatbots. Meanwhile, the highest-leverage automation in American healthcare is already deployed, already at scale, and pointed in exactly one direction: saying no, faster.

The asymmetry is the whole story

Start with how often "no" happens. KFF's analysis of federal data found that insurers on HealthCare.gov denied roughly one in five in-network claims in 2024. That rate has hovered near 19 percent for several years, and it ranges from about 3 percent to 36 percent depending on the insurer. Eighty-five million in-network claims denied, in a single year, in a single slice of the market.

Now look at what happens next. Fewer than one percent of denied claims are ever appealed. And here is the part that should bother you: when patients do appeal, they win a meaningful share of the time. In 2024, insurers reversed roughly a third of the internal appeals they received. Other years it has been closer to half.

Read those two facts together. A large fraction of denials do not survive scrutiny, but almost no one applies scrutiny. The denial is cheap to issue and expensive to fight, so most of them are never fought. That gap, between how easy it is to deny and how hard it is to contest, is the entire business model. AI did not create it. AI just widened it.

When the algorithm is the adjudicator

The Cigna story is about volume. The next one is about judgment.

UnitedHealth's subsidiary naviHealth used an algorithm called nH Predict to estimate how long a patient would need post-acute care, the nursing-home and rehab stays that follow a hospitalization. According to a class-action complaint filed in late 2023, the company used those predictions to cut off coverage, and managers were pressured to keep patients' actual stays within one percent of the algorithm's projection. The lawsuit alleges that roughly 90 percent of the algorithm's denials were reversed on appeal.

If that allegation holds, think about what it means. A 90 percent reversal rate is not a tool that is occasionally wrong. It is a tool that is wrong almost every time it is challenged, and it stays in service precisely because it is so rarely challenged. The reversals only happen for the sliver of patients and families with the energy to fight. Everyone else simply loses their coverage.

This is not a fringe case. In October 2024, the Senate Permanent Subcommittee on Investigations published a report, built on more than 280,000 pages of internal documents, on how the three largest Medicare Advantage insurers handle prior authorization for post-acute care. It found that UnitedHealthcare's denial rate for that category climbed from 10.9 percent in 2020 to 22.7 percent in 2022. The report tied the increase directly to the insurers' adoption of predictive technologies. The denials did not rise because patients got healthier. They rose because the tooling got better at saying no.

Why this is the AI story that matters

I think the reason this gets less attention than it should is that it does not look like the AI we were promised. There is no chat interface, no demo, no founder on stage. It is a batch job. It runs quietly, at the seam between a claim and a payment, and its output is the absence of a check that would otherwise have been written.

But it is, functionally, one of the most consequential deployments of machine decision-making in any industry. It operates with minimal human review, it acts on millions of people, and the cost of its errors lands on patients and on the providers who treated them in good faith.

The uncomfortable thing about the asymmetry is that it is stable. As long as denying is near-free and appealing is costly, the rational move for a payer is to deny aggressively and let the appeal rate stay under one percent. Regulation is starting to push back, and I will write about that next. But the underlying physics has not changed: the side that automates "no" is operating at a scale the other side cannot match by hand.

The interesting question, and the one I keep coming back to, is what happens when the other side stops trying to match it by hand.