Claims used to crawl through 21 days of manual review. We rebuilt the pipeline around an LLM extraction core and cut it to 4, with 70% of claims never touching a human.
Every claim landed in a shared queue where adjusters manually keyed fields from PDFs, scanned forms, and email attachments before any decision could begin. Triage alone consumed days, and identical low-risk claims waited behind complex ones.
The cost wasn't just time. Inconsistent data entry made fraud screening shallow and audits painful, since the record of why a claim was approved lived in scattered email threads.
We put an LLM extraction core at the front of the pipeline. OpenAI reads each document into a strict JSON schema, validated against the policy on file, so downstream rules operate on clean structured data instead of raw text.
A PostgreSQL rules engine then scores coverage and fraud in one pass. High-confidence claims auto-adjudicate; everything else routes to adjusters with the model's extracted fields and reasoning pre-loaded, turning hours of triage into a single confirm-or-correct screen.
| Dimension | Before | After |
|---|---|---|
| Avg cycle time | 21 days | 4 days |
| Manual touch rate | 100% | 30% |
| Field extraction | Adjuster keying | LLM + schema validation |
| Fraud screening | Spot-checked | Every claim scored |
| Adjuster load | Full triage | Exceptions only |
| Audit trail | Email threads | Immutable event log |
Every LLM output is checked against a strict JSON schema and the customer's policy before it advances.
Claims clear automatically only when extraction and rules confidence exceed a tuned threshold.
A rules engine scores every claim for anomalies instead of relying on random spot checks.
Reviewers see extracted fields, model reasoning, and source documents side by side for fast decisions.
FastAPI and S3 handle OCR and extraction asynchronously so intake never blocks on processing.
Every decision, override, and payout writes to a tamper-evident event trail for compliance.
“"We didn't replace adjusters, we deleted the 17 days of busywork standing between them and the decision."”
The platform turned a 21-day manual slog into a 4-day, mostly-automated flow: 70% of claims now settle without a human touch, and the adjusters who remain spend their time on genuine exceptions instead of data entry. Faster, more consistent decisions pushed CSAT up 45%, while per-claim fraud scoring and an immutable audit trail made the whole operation more defensible than the email-thread era it replaced. Built on Python, FastAPI, OpenAI, PostgreSQL, and AWS, it scales with claim volume instead of headcount.