A mid-market lender drowning in manual claims review deployed a document-intelligence pipeline that cut handling time by 71% — with a human still signing off on every decision.
The problem
Claims arrived as a mess of PDFs, scans and emails. Analysts spent most of the day keying data and cross-checking documents, and turnaround had stretched to days while the backlog grew.
What we built
A pipeline that classifies each incoming document, extracts the fields that matter, validates them against policy rules, and surfaces a pre-filled, flagged summary for an analyst to approve. The model handles the typing; the human keeps the judgment.
The outcome
Average handling time fell 71%, the backlog cleared inside a quarter, and accuracy improved because analysts reviewed exceptions instead of re-typing every case. Throughput roughly tripled with the same headcount.
Why it stuck
Evals on a labeled claim set gate every model update, and a confidence threshold routes anything uncertain straight to a person. Reliability — not raw automation — is what earned the team's trust.