The arithmetic is startling, if somewhat depressing. In 2024, healthcare providers filed appeals on just 11.5% of the 4.1 million insurance claims that got denied. Yet when they did bother to fight back, they won 80.7% of the time.
Do the math on that gap—the difference between what providers could theoretically recover and what they actually do—and you arrive at a figure in the billions. It's the kind of inefficiency that makes Silicon Valley sit up and take notice. Enter Insurf, a San Francisco startup backed by Y Combinator that's now taking aim at this sprawl of abandoned revenue with a product called Inveto.
The company's pitch is straightforward: automate the grunt work of turning a denied claim into a properly documented, physician-approved appeal. But the method matters. Unlike some competitors in the burgeoning field of AI-powered denial management, Insurf has made an unusual design choice. Its system extracts facts from medical records and binds them to exact source text. It doesn't paraphrase. It doesn't synthesize. And crucially, it doesn't invent.
"Citation-gap and quote-not-found are hard blockers," the company's documentation states, almost defiantly. If the clinical evidence doesn't back up a claim in the appeal, the software won't let you submit it. No papering over inconvenient gaps.
Upload, Extract, Anchor
Inveto's workflow is built around constraints. Providers upload denial letters, clinical notes pulled from the electronic health record, and snapshots of payer policies. Then the system—powered by OpenAI's frontier models—does one thing well: it pulls facts from the medical chart and tethers each one to its verbatim source.
The process unfolds in six phases: intake, ground, match, assemble, review, track. During the "ground" stage, every material fact gets anchored to exact source text. Hover over a claim in the draft appeal, according to demo screenshots, and you'll see the precise clinical note, policy clause, or medical compendium entry it came from. Three sources, zero fabrications.
A physician still has to review and sign off on the assembled packet before it goes anywhere. Insurf isn't trying to replace clinical judgment—at least not yet.
Behind this citation-first interface, the company is quietly constructing something more ambitious: what it calls a "Coverage-Decision Graph Transformer." Each resolved appeal flows back into this internal model, which tracks how payers behave, what precedents matter, which evidence tends to sway decisions. The graph doesn't write appeals itself. It recommends next moves—whether to appeal at all, which clinical details to foreground, what the expected recovery value might be.
That same decision graph will eventually power a second product, Surely, aimed at patients shopping for health insurance. Surely simulates what a year of healthcare might actually cost by layering in premiums, deductibles, coinsurance, network quirks, and—here's the twist—denial risk. It's a consumer-facing slice of the same underlying data engine.
A Crowded Field, Perhaps
Insurf is hardly alone here. Aegis, Incerto, InvisaClaim, Overturn, Appeal AI—there's no shortage of startups chasing the same pile of rejected claims. What distinguishes them is methodology. Some lean hard into full automation. Others deploy monitoring agents. Still others draft appeals but skip the citation rigor that Insurf emphasizes, for better or worse.
The regulatory backdrop is shifting in ways that might help. CMS rules requiring FHIR-based APIs for prior authorization decisions are being gradually implemented through 2026-2027, with tighter timelines—72 hours for expedited reviews, seven days for standard ones. Separately, updated health IT certification criteria now include standards for electronic prior authorization. Insurf integrates via SMART on FHIR to read clinical charts and tracks outcomes through billing files and, yes, inbound fax. Because healthcare infrastructure is nothing if not a patchwork.
The burden isn't trivial. A survey released by the American Medical Association found that physicians and their staff burn roughly 13 hours a week navigating prior authorization. Ninety-five percent say it delays necessary care. A quarter report a serious adverse event they attribute to prior-auth delays. Those numbers tend to come up in investor pitches.
Young Company, Old Problem
Insurf's founders—Chang Lu and Bryan Chung, both Brown graduates—incorporated in California in mid-2026, according to available filings, though some details in earlier company materials appear potentially outdated. Lu studied neuroscience and economics and worked on zero-shot modeling research at Emory's Winship Cancer Institute. Chung competed internationally in physics, reached the USACO Platinum level in computer science competitions, and has two IEEE publications to his name.
The product is in what the company calls "private pilot readiness," with production handling of protected health information still disabled pending security and compliance sign-offs. There's no SOC 2 report yet, though Insurf is working toward one via Vanta, with AWS, Vercel, Neon, and OpenAI as covered vendors. Pricing isn't public. Instead, the company is framing early access around "Founding Data Partnerships"—90-day pilot evaluations with a benchmark of eight to 12 appeals processed in the first month.
Insurf has promised to publish denominators once pilots mature: drafts generated, physician attestations secured, appeals resolved, overturns achieved. Transparency as product strategy, or at least marketing.
For now, the product is live in private beta. The regulatory environment is inching toward transparency and interoperability, however haltingly. And those billions in recoverable revenue? Still sitting there, waiting for someone—physician, software, or some hybrid of the two—to go retrieve them.
