LLM prior-auth review,
cut from days to minutes.

The challenge: a manual prior-auth backlog.

Prior authorization is one of the slowest, most manual workflows in US healthcare. Our client, a Series B platform serving specialty clinics, had a queue of mixed-format insurance documents that clinical operations staff read by hand. The backlog kept growing, and hiring more reviewers wasn't going to fix it.

We see the same pattern across our healthcare AI solutions work: high volume, strict rules, expensive human attention.

We started with one hypothesis: most cases follow predictable patterns and can be auto-routed, and the rest need human judgment. Our job was to find that boundary, measure it and automate everything on the safe side of it.

The approach: LLM evals before models.

We opened with a discovery sprint: we read 200 anonymized cases and wrote the eval criteria by hand before any code existed. Those criteria became the test set. Next came a thin prototype (a small classifier plus an LLM extractor) run against the client's real document corpus.

No model selection until we had an evaluation harness. No scaling until the eval passed, reproducibly.

From there our LLM application development team built the production pipeline: structured output, confidence-banded routing, and a human-in-the-loop interface for cases the model isn't sure about. Every decision is logged for audit and retraining.

The system: a HIPAA-compliant LLM pipeline.

Documents arrive from clinics through three channels: fax, secure email and portal upload. Each one is parsed into a canonical format, classified, then either auto-routed or sent to a clinical reviewer with the model's reasoning attached. Reviewers can correct the model, and every correction feeds a continuous-evaluation loop we re-run weekly.

Compliance is part of the architecture, not a layer on top:

• End-to-end encryption and role-based access control
• BAA-eligible infrastructure in a HIPAA-compliant AWS account
• A complete audit log, indexed by case ID
• Isolated test and production environments, defined in reviewable infrastructure-as-code

LLM prior-auth automation results at GA.

We measured against the eval set agreed in discovery. Routine cases now clear in minutes instead of days, and reviewers spend their time on the cases that need clinical judgment.

What we'd do again.

Building the eval harness in week one paid for itself five times over. Every time we tried a new model or prompt strategy, we knew within 10 minutes whether it was better. There was no debate. There was a number. It's now standard in our LLM evaluation and QA practice.

We also kept the human-in-the-loop interface deliberately generous. Reviewers see the full document, the model's reasoning and prior decisions on similar cases, and can override anything in two clicks.

What's next.

The system is in steady-state operation. We co-own it with the client's internal team, and one of our engineers still spends a day a week on tuning, eval review and incident response. The eval set now holds over 1,400 cases and is growing.