Automated Claims Processing with Document AI

The Challenge

A property and casualty insurer processing 15,000 claims per month was handling every claim manually. Claims arrived as bundles of scanned documents — police reports, medical bills, repair estimates, photos — that were printed, sorted, and keyed into the claims system by a team of 60 data entry clerks. The average claim took 12 days from first notice of loss (FNOL) to payment, and data entry errors caused 8% of claims to be reworked.

Our Approach

We built a document AI pipeline that automatically extracts, validates, and routes claim documents, enabling 80% of straightforward claims to be processed without human intervention.

Document Classification: Incoming documents are classified into 12 types (police report, medical bill, repair estimate, photo evidence, correspondence, etc.) using a fine-tuned document classification model. Classification accuracy is 97.5%.

Data Extraction: Each document type has a custom-trained extraction model built on Google Document AI. The models extract structured fields — for medical bills: provider name, CPT codes, billed amounts, dates of service; for repair estimates: line items, labour hours, parts costs, total estimate; for police reports: incident date, parties involved, officer narrative, report number.

Validation Engine: Extracted data passes through a rules engine that checks for consistency (does the incident date on the police report match the FNOL date?), completeness (are all required document types present?), and fraud indicators (is the repair estimate from a shop on the watch list? does the medical billing exceed typical costs for the diagnosis code by more than 2 standard deviations?).

Routing: Claims that pass all validation rules are auto-adjudicated and routed to payment. Claims with validation failures or fraud flags are routed to the appropriate human reviewer (adjuster, special investigations unit, medical reviewer) with the extracted data pre-populated and the specific issue highlighted.

Results