# Claim Triage

> An insurance review workspace linking claim documents, policy and repair checks, and traceable recommendations for a human officer.

Trace the evidence. Support the officer's decision.

- URL: https://weaveward.com/work/motor-insurance-claim-triage
- Stage: Prototype
- Year: 2026
- Industry: Insurance
- Context: Internal product
- Capabilities: AI solutions, Computer vision, Business systems
- Technologies: Python, FastAPI, PostgreSQL + pgvector, RapidOCR, GoRules ZEN, OpenCV, RF-DETR-Seg adapter, React, TypeScript

## The challenge

A motor claim can arrive as a mixture of phone photos, scanned identity documents, a policy, a workshop estimate and a multi-page PDF. An officer must check completeness, compare details across documents, assess repair costs and apply authority rules before making a decision.

The hard part is keeping the reasoning visible. A recommendation is only useful when the officer can see what supports it, what is missing and what needs another person's review.

## What we built

Claim Triage brings intake, document processing and review into one workspace.

- **Guided intake.** Policyholders and officers upload documents through a checklist with photo guidance and instant quality checks. The system accepts PDF, JPG, PNG and HEIC files, with up to 50 files per claim
- **Evidence the officer can inspect.** OCR, page classification and PDF splitting organise the bundle. Extracted values retain their confidence, page and position, so the workbench can highlight the source
- **Cross-checks and repair estimates.** Policy coverage, identity, vehicle details and invoice arithmetic are checked together. A damage-model adapter feeds severity rules, parts prices and labour rates into an editable cost range
- **Review and approval.** Officers receive reasons, missing-information requests, risk signals and proposed next steps. Supervisors, investigators and surveyors have their own queues, with explicit authority limits and decision records
- **Rules operations can change.** Versioned decision tables control coverage, deductibles, red flags and routing. Signed webhooks and a core-system adapter support integration

## How it works

### Manual review and approval

The processing pipeline records each stage and hands failures to manual review. A fraud signal can trigger an investigation; it cannot automatically decline a claim. Declines and decisions above an officer's authority require additional approval.

### Optional, privacy-gated language models

Language-model features are optional, sit behind a privacy gate and fall back to rules or templates. The review workflow remains useful without a language model.

## Current stage

This is a working internal prototype tested with synthetic claims. The historical evaluation recorded 163 of 163 key fields correctly extracted across 50 synthetic cases using OCR and rules. The documented laptop timing is a development result, not a production service guarantee.

Real damage assessment needs weights trained and validated for the insurer's photos. The demo uses annotated synthetic images; real claims without a trained model are sent for manual or surveyor assessment. No live insurer pilot or production deployment is claimed.

## Where this fits

The same pattern applies to document-heavy decisions: organise the evidence, automate repeatable checks and keep human authority clear. Explore our [AI solutions](/services/ai-solutions), [computer vision](/services/computer-vision) and [business systems](/services/business-systems), or [discuss your review workflow](/contact).