Every year, 1 in 7 insured Americans has a health insurance claim denied.
The denial letter isn't the hard part. What comes after is: someone, usually a medical billing specialist, not a doctor, not a lawyer, has to read the denial, decode the reason code, dig through payer policy, track down clinical evidence, draft an appeal, check every citation by hand, watch the deadline, and get it submitted before the window closes.
That's hours of procedural work standing between a patient and care they were already supposed to get.
Where should an AI agent stop in an enterprise workflow?
Not "how autonomous can I make this." A narrower, harder question: where does preparation end and a human decision begin?
The answer I landed on became the whole architecture:
AI prepares. AI verifies. Human authorizes.
That boundary, not the model and not the agent count, is the actual product. Everything else exists to make that boundary enforceable.
DenialDefender is an 8-agent Google ADK fleet that runs an insurance denial through a governed pipeline instead of a single prompt.
| Agent | Job |
|---|---|
| Patient Advocate | Intake, urgency, deadline extraction |
| Denial Triage | Reason-code classification, appealability assessment (Gemini 3.6 Flash) |
| Policy Research | Payer policy contradiction retrieval |
| Evidence Assembly | Clinical evidence with provenance tiers |
| Citation Verification | Citation resolution and SHA-256 hash matching |
| Letter Drafting | Evidence-grounded appeal letter with inline citations (Gemini 3.6 Flash) |
| Quality Review | Adversarial 7-check battery that refuses to pass until every claim is grounded |
| Deadline Tracker | Deadline monitoring and escalation |
Two human-in-the-loop gates sit inside that pipeline. Gate 1: a human confirms the triage before research begins. Gate 2: a human approves the final letter before it's submitted. No letter reaches a payer without explicit human sign-off.
Two agents in the fleet do the real reasoning, and both run on Gemini 3.6 Flash via the Vertex AI global endpoint. Denial Triage classifies the denial and estimates appeal success. Letter Drafting generates the appeal letter itself, with citations that resolve back to hashed evidence. Take away Gemini 3.6 Flash and those two agents have nothing to reason with.
Trust isn't a claim here. It's four mechanisms, each with an auditable proof:
I built three evaluation harnesses instead of just asserting the system works.
Honesty is an asset, not a liability.
The outcome-learning dataset uses 5 public records and 45 synthetic ones. I labeled it that way instead of presenting all 50 as real. The GEAP components report backend=local
honestly instead of hiding it. Every one of those choices made the architecture stronger, because the claim can be independently checked against the underlying evidence.
AI prepares. AI verifies. Human authorizes.