Cutting claims cycle time with document AI and automated triage

The insurer received first notice of loss across five channels — a customer portal, a broker portal, email, fax, and a call centre — and every one of them ended the same way: a person reading a document and retyping its contents into the claims system.

Roughly seven in ten inbound claims were low-complexity and highly repetitive. They still consumed the same intake effort as a complex liability claim, because triage happened after data entry rather than before it.

The consequences showed up in cycle time and in retention. Average time from first notice of loss to adjuster assignment sat above four days in peak weeks, and the claims that suffered most were the small, straightforward ones where a policyholder is most sensitive to how long it takes.

Adjusters were the scarcest resource in the business and were spending a material share of their week on transcription. Meanwhile fraud referral depended on a rules engine that only saw structured fields, so signals sitting in the unstructured documents were never evaluated.

A previous vendor proposal had recommended replacing the policy administration system. Leadership rejected it. The system worked; the intake process wrapped around it did not.

The objectives were set accordingly: automate document capture across all intake channels, triage claims before human handling, route low-complexity claims straight through, surface fraud signals from unstructured content, and do all of it without modifying the policy administration system.

The work was scoped as business process automation with machine learning inside it, not as an AI project. The sequencing followed the data: capture first, then classification, then routing, then straight-through processing. Each stage had to be reliable before the next one was allowed to depend on it.

Document capture was built as an intake service that normalised every channel into one pipeline. Optical character recognition handled scanned and photographed documents; a layout-aware extraction model pulled policy numbers, loss dates, claimant details, and loss descriptions into a structured claim object with a confidence score attached to every field.

Confidence scores drove the automation boundary. Fields above threshold passed through untouched. Fields below threshold were routed to a human review queue that presented the document and the proposed value side by side, so correction took seconds rather than requiring a full re-read. Every correction became training data.

A classification model then assigned claim type, complexity band, and an initial fraud signal score using both structured fields and the unstructured loss description. Low-complexity claims with clean extraction and no fraud signal were created directly in the policy administration system through its existing integration interface and assigned automatically.

Everything else was routed to an adjuster with the extracted data pre-populated and the source documents attached. Even where automation did not complete the claim, it removed the transcription step.

Adjusters retained override authority at every point, and every automated decision was logged with the model version, input, confidence, and outcome — a requirement from the compliance function before any model was allowed near a live claim.