Physicians complete an average of 39 prior authorization requests per week, spending about 13 hours a week on the process, according to the AMA's 2024 Prior Authorization Physician Survey of 1,000 practicing physicians. 94% say prior authorization contributes to burnout, and 93% say it delays patient care. Meanwhile, KFF's analysis of CMS data found that 80.7% of Medicare Advantage prior authorization denials that were appealed in 2024 were fully or partially overturned, yet only 11.5% of denials are ever appealed at all. The evidence to win most appeals already exists. Almost nobody has the hours to go find it. Denial Appeal Agent closes that gap. It reads a denied insurance prior authorization request (the denial letter, the patient's chart, and the payer's own policy) and finds the evidence, scattered across all three documents, that defeats each denial reason. It then drafts a fully cited appeal letter for a human clinician to review and send. The core pattern behind most winnable denials: insurers often cite a rule while overlooking their own documented exception to it. A denial might claim there is no completed 3 month drug trial on record, while the chart shows the drug was stopped early for documented toxicity, which satisfies a specific policy exception the denial never mentions. Finding that requires cross referencing three unstructured documents by hand. The agent does it live, with every tool call visible in the UI. Built as a raw Anthropic SDK tool use loop, with no LangChain and no framework. Four core tools (parse denial, search chart, check policy clause or exception, draft appeal) that the model calls in whatever order it decides, not a hardcoded pipeline. On top of that, an adversarial reviewer agent using Claude Haiku critiques each draft like a skeptical payer reviewer and can force a revision before the letter is finalized. This is an evaluator optimizer loop, combined with model routing (Sonnet drafts, Haiku reviews) and parallel execution of independent tool calls. Tested against both a synthetic demo case and a real, unedited UnitedHealthcare policy PDF downloaded during the event, proving the search tooling generalizes beyond a cherry picked sample. No autonomous medical judgment involved. The system finds and cites evidence, and a human decides whether to send it.
Team
Placement
Participant
