A property and casualty insurance company processed 8,000 claims per month with an average handling time of 5.2 days from first notice to initial assessment. The claims team was understaffed, and seasonal events (storms, floods) created backlogs that pushed handling times to weeks. The company deployed AI to automate intake, initial assessment, and routing.

The Architecture

The system handles the claims lifecycle from first notice through initial assessment and adjuster assignment.

Intake processing - Claims arrive via web form, mobile app, email, and phone (transcribed). Each claim includes a description of the loss, policy number, date of loss, and supporting documentation (photos, police reports, repair estimates). The intake layer normalizes all channels into a standard claim record and triggers document processing for attachments.

Document extraction - Supporting documents are processed through a document AI pipeline. Photos are analyzed for damage assessment using computer vision (Amazon Rekognition for object detection, a multi-modal LLM for damage severity assessment). Written documents (police reports, medical records, repair estimates) are processed through OCR and entity extraction to pull key data: amounts, dates, parties involved, and injury descriptions.

Claim assessment - An LLM evaluates the complete claim package: the claim description, extracted document data, policy coverage details, and historical claim patterns for similar losses. The model produces: coverage determination (covered, not covered, requires review), estimated loss amount range, complexity classification (simple, moderate, complex), and fraud risk indicators.

Routing and assignment - Simple claims with clear coverage, low estimated amounts, and no fraud indicators route to auto-adjudication for fast settlement. Complex claims route to experienced adjusters with the full AI assessment as a starting brief. Claims with fraud indicators route to the special investigations unit.

Key Lessons

Photo analysis was surprisingly effective for property claims. The system could assess hail damage severity from roof photos with 82% agreement with field adjuster assessments. This enabled remote assessment for many claims that previously required an in-person inspection, reducing cycle time significantly for straightforward damage claims.

Coverage determination required extreme caution. The system could correctly identify coverage for 90% of claims, but the 10% it got wrong created significant problems - incorrectly denying a legitimate claim or incorrectly approving a non-covered loss. The team positioned AI coverage determination as a recommendation that required adjuster confirmation for all denial decisions and for approval decisions above $10,000.

Seasonal surge was the real test. The system performed well under normal volume, but its value was proven during a major hailstorm that generated 3x normal claim volume in one week. The AI system processed initial intake and assessment for 70% of surge claims without human involvement, preventing the backlog that would have taken months to clear manually.

Fraud detection benefited from cross-claim analysis. Individual claims that passed fraud checks were sometimes part of organized fraud rings. Cross-claim analysis - looking for patterns across multiple claims from the same geographic area, same contractors, or similar damage descriptions - identified fraud rings that single-claim analysis missed.

Results

Average handling time from first notice to initial assessment dropped from 5.2 days to 18 hours. Straight-through processing rate for simple claims reached 45%. Customer satisfaction scores improved by 22% due to faster response times. Claims processing costs decreased by 30% through a combination of automation and better resource allocation.