Challenge
The client - a mid-tier insurance group with 400k active policies - processed claims almost entirely by hand. Adjusters manually keyed data from scanned documents (medical reports, invoices, police reports) into a legacy system. Average claim processing took 14 business days. Document misclassification and data entry errors caused 23% of claims to require rework. During storm seasons, the backlog grew faster than staff could handle.
Solution
We built an end-to-end claims processing pipeline that automates the repetitive parts and routes complex cases to human reviewers:
- Document intake via Azure AI Document Intelligence - OCR with field extraction for 18 document types (medical, automotive, property)
- Classification model that routes claims to the correct processing track with 96.3% accuracy
- Rules engine for straight-through processing of simple claims (broken windshield, minor water damage, standard medical reimbursement)
- Reviewer portal (React/TypeScript) showing extracted data side-by-side with source documents, confidence scores, and suggested decisions
- Audit trail with full provenance - every automated decision traceable back to source data and rule version
- Queue management with priority scoring based on claim value, policy type, and SLA deadline
Results
- 78% of claims now auto-processed without human intervention (up from 0%)
- Average decision time dropped from 14 days to 3 minutes for auto-processed claims
- 12x throughput increase during peak periods with no additional headcount
- Rework rate down from 23% to 4.1%