Fordel StudiosThis case study describes a real engagement. Client identity, proprietary details, and specific metrics are anonymized or approximated under NDA.
Claims Processing Automation for Motor Insurance
Manual claims processing averaging 5 business days from submission to adjuster assignment. High variability in damage assessment categorization — similar damage types were being routed to different adjuster tiers, creating inconsistent settlement timelines and adjuster workload imbalances.
Document ingestion pipeline combined with image analysis for damage classification, producing structured claims records with automated severity scoring and adjuster routing logic.
This engagement automated the intake and initial assessment layer of a motor insurance claims operation. The system handles the full intake workflow: document parsing for policy documents, incident reports, and repair estimates; image analysis for vehicle damage classification; severity scoring against configurable routing rules; and automated assignment to the appropriate adjuster queue. The pipeline was built to integrate with the existing claims management system via API rather than replacing it, which constrained the output schema but simplified deployment. The system processes claims end-to-end in under 2 minutes from document upload to adjuster assignment queue entry.
The Challenge
Motor insurance damage images vary significantly in quality — many are taken by policyholders on mobile phones under poor lighting conditions, at odd angles, or with obstructions. The classification model had to handle this noise without requiring standardized photography, which would have created a policyholder friction problem. Document parsing complexity was also significant: policy documents arrived in multiple formats (PDF, scanned images, and occasionally faxed paper scans that had been photographed) with no standardized layout across the vehicle makes and policy types in the portfolio. A secondary constraint was the existing claims management system, which had been in operation for over a decade and had no modern API — integration required a custom adapter layer built against a SOAP-based interface.
How We Built It
Data audit and damage taxonomy (Weeks 1–2): We reviewed 2,000 closed claims with known outcomes to map the damage classification taxonomy in use by the adjuster team. This produced 14 damage severity categories across four vehicle zones, which became the classification target for the image analysis model. We also identified the 8 document types that appear in 95% of claims, establishing the parsing targets for the document pipeline.
Document ingestion pipeline (Weeks 3–5): The ingestion pipeline handles PDF, scanned image, and photographed document inputs using a pre-processing pass (deskew, contrast normalization, resolution upscaling) before OCR. FastAPI serves as the ingestion endpoint, with AWS S3 for raw asset storage and PostgreSQL for extracted structured data. Field extraction for each document type is handled by Anthropic Claude with per-type extraction prompts, returning structured JSON that maps to the claims management system schema. Validation rules catch extraction anomalies before they reach downstream systems.
Image analysis and damage classification (Weeks 6–8): We fine-tuned OpenAI's Vision model on the 2,000-claim image dataset, with augmentation (rotation, brightness variation, partial occlusion) to address the quality variability in real-world mobile photography. The model classifies damage zone (front, rear, side, roof) and severity tier for each submitted image, aggregates across multiple photos per claim, and produces a single composite severity score with per-zone breakdown. Confidence thresholds route low-confidence assessments to human review rather than automated adjuster assignment.
Routing logic and integration (Weeks 9–10): Adjuster routing rules were configured in collaboration with the operations team, mapping severity score ranges to adjuster tiers with override logic for high-value policies and repeat claims. The SOAP adapter layer translates the pipeline output into the legacy claims management system's expected format. A Redis queue buffers claim packets during peak intake periods to handle volume spikes without processing delays. End-to-end processing time from document upload to queue entry averages 1.2 minutes at baseline load.
What We Delivered
Average claim processing time from document submission to adjuster assignment dropped from 5 business days to 1.2 days — a 72% reduction. The document ingestion pipeline processes incoming claims within 90 seconds of submission during standard operating hours. Adjuster assignment no longer requires a manual triage step, freeing the operations team from the intake queue entirely.
Damage classification accuracy reached 94% against the ground-truth taxonomy used by senior adjusters, measured on a held-out test set of 400 claims. The previous manual categorization had an inter-rater reliability of approximately 76%, meaning the automated system is more consistent than the manual process it replaced. Claims routed to the wrong adjuster tier dropped from approximately 18% to under 4%.
The routing consistency change has produced measurable secondary effects. Adjuster workload distribution improved — the previous manual triage was inadvertently routing a disproportionate share of complex claims to a subset of senior adjusters, creating a bottleneck. Automated routing distributes load according to the configured rules, and average adjuster queue depth is now consistent across tiers throughout the business day.
Ready to build something like this?
Tell us what you are building. We will scope it, price it honestly, and give you a clear plan.
Start a ConversationFree 30-minute scoping call. No obligation.