AI-Assisted Document Processing for Insurance Claims

The situation

Insurance claims run on documents: first notice of loss forms, repair estimates, invoices, police reports, medical records and photos. Traditional OCR with fixed templates handles the tidy ones and fails on everything else. Adjusters end up reading each document, retyping values into the claims system and flipping to the policy wording to check coverage, limits and exclusions.

Large language models are good at reading messy documents, but teams in this situation cannot simply hand claims to a model. Outputs can be wrong in confident-sounding ways, personal data needs strict handling, and regulators expect every decision to be explainable. The goal is to take the retyping off adjusters, not to take adjusters out of the process.

Our approach

1. Start with a labeled evaluation set

Before building anything, we work with claims staff to collect a representative sample of real documents, anonymized where required, and label the fields that matter. This evaluation set defines what “good” means for each document type and field. Every change to prompts, models or OCR settings is measured against it, so improvements are proven rather than assumed.

2. Capture, classify and protect documents

Documents from email, the customer portal and the scanning room land in encrypted object storage with retention rules. OCR extracts text and layout, and a classifier identifies the document type. Personal and health information is detected and tagged, and only the minimum needed for a task is passed to the model. Model calls run inside the cloud account through a managed service, with no data used for training and with access logged.

3. Extract fields, grounded in the policy

An LLM extracts fields into a strict JSON schema for each document type, with the page and region each value came from. For coverage-related fields, retrieval-augmented generation pulls the relevant clauses from the claimant’s policy documents, indexed in pgvector, so the model reasons over the actual wording instead of general knowledge. The output cites which clauses were used.

4. Keep people in the loop

Validation rules cross-check extracted values: dates within the policy period, totals that add up, policy numbers that exist. A confidence router combines model signals and rule results against thresholds set per field. High-confidence fields are pre-filled; low-confidence or high-impact ones, such as coverage decisions and large amounts, always go to the review queue. Adjusters see the source document next to the extracted data, confirm or correct each value, and their corrections feed back into the evaluation set.

5. Integrate and audit everything

Only reviewed and approved data is written to the claims system through its API. The audit trail records the source document, OCR output, model and prompt version, retrieved clauses, confidence scores and the person who approved each value. Auditors can reconstruct how any field was produced.

How we deliver it

We usually begin with an IT consulting engagement to agree on scope, data handling, model options and success criteria with claims, IT, legal and compliance. A pilot then covers one or two high-volume document types in shadow mode: the system extracts, adjusters work as usual, and we compare. Only when accuracy on the evaluation set and in shadow mode meets the agreed bar do we switch on pre-filling, one document type at a time. Workflows run on Temporal so long-running reviews and retries are reliable, and infrastructure is defined in Terraform.

Is this relevant to you?

If adjusters spend their days retyping documents, or earlier OCR projects stalled on format variety, this approach is a good fit. Explore our custom software development service or talk to us about your claims workflow.