At FurtherAI, we tackle one of insurance’s thorniest challenges: extracting structured data from chaotic loss runs. These documents—akin to business credit reports—track claims history across hundreds of insurers, each with wildly inconsistent formats. One 180-page document might split 10 claims across four tables, while another buries critical metadata in hidden headers. Traditional extraction tools fail here, but our self-correcting AI agent now achieves 95% row count accuracy by combining multiple extraction backends with autonomous validation.

Our breakthrough came shifting from rigid prompt engineering to a feedback-driven system. The agent independently calls extraction APIs, cross-references document totals, and re-examines suspicious sections using page-specific tools. For a 5-page loss run with 35 claims, it detected duplicate entries in Commercial Auto sections by visually inspecting pages 2-3. This approach outperformed weeks of prompt tuning, reducing errors caused by ambiguous formatting like subtotal rows masquerading as claims.

The agent’s three tools—extract_claims, focus_pages, and validate_totals—enable iterative debugging. When totals mismatch (like 38 extracted vs 35 stated claims), it zooms into Workers Comp and Commercial Auto sections to identify duplicates. Even $0 financial fields are decoded contextually: closed claims versus placeholder data. This reasoning mirrors human pattern recognition, handling edge cases like summary rows mistaken for individual claims.

By treating extraction as a validation loop rather than a single-pass task, we address the core problem: document semantics, not just text. The system adapts to carriers’ unique conventions, from blank cells meaning "same as above" to policy numbers hidden in section headers. This isn’t just about better OCR—it’s about building AI that understands insurance documents the way humans do.