Claims and Underwriting Documents on AWS: Textract, Bedrock Data Automation, or an LLM?

First notice of loss forms, medical records, police reports, loss runs, invoices and supplemental applications arrive as PDFs, scans and photos. Three ways to turn them into data on AWS, with current per-page prices, the accuracy you should measure, and the human review loop that keeps it safe.

Insurance runs on documents that someone retypes. Automating that is one of the best-understood uses of AI in the industry, and the technology choice matters less than the process around it.

Three approaches

The pipeline

1. Classify each incoming document: claim form, medical record, estimate, police report, invoice.

2. Extract with the cheapest method that is accurate for that class.

3. Validate against the systems of record: does the policy number exist, does the date of loss fall in the policy period, do line items sum.

4. Route anything below a confidence threshold, or failing validation, to a person, with the source highlighted.

5. Record what was extracted, by which version, and who corrected it. Corrections become your next test set.

Measure before you automate

Build a labelled set of a few hundred real documents per class and measure field-level accuracy, not a demo. Set thresholds per field: a misread claimant phone number is a nuisance, a misread loss amount is a payment error. Start in a mode where people review everything and the system only pre-fills, then lower review as measured accuracy justifies it.