From 50-Page Submissions to Instant Risk Scores

Underwriters spend their days buried in submissions: 50-page engineering reports, scanned loss runs, handwritten field notes, financial statements with embedded tables. 64% of insurers cite document processing as their top AI priority, and for good reason.

The document problem is multimodal

Insurance submissions are not clean PDFs. They include scanned pages, photographs, handwritten annotations, tables that span multiple pages, and inconsistent formatting across brokers.

ai_parse_document solves this natively

Databricks ai_parse_document is a built-in AI Function that extracts structured content from unstructured documents. It handles PDFs, images (JPG, PNG), DOCX, and PPTX files. It captures layout information, parsed tables, bounding boxes, figures, and comprehensive document structure, all directly in SQL or PySpark. No separate OCR pipeline needed.

The underwriting copilot uses three agents:

1. Document Intelligence Agent: Uses ai_parse_document to extract text and structure, then ai_extract to pull COPE data (Construction, Occupancy, Protection, Exposure) and risk factors
2. Risk Scoring Agent: Cross-references extracted data with loss history and market benchmarks via Feature Store
3. Recommendation Agent: Uses ai_gen to suggest policy terms and pricing, flagging anomalies for underwriter attention

Databricks-native architecture

• Lakeflow Connect syncs loss history from enterprise applications via managed connectors
• Spark Declarative Pipelines handle submission processing ETL (streaming tables for incoming submissions, materialized views for enriched profiles)
• Lakeflow Jobs orchestrate the end-to-end underwriting workflow
• Delta Lake stores submissions and loss data
• Feature Store serves risk scoring features with point-in-time correctness
• Unity Catalog governs everything: data, models, functions, pipelines, and serving endpoints
• All models are accessed via AI Gateway on Databricks Model Serving
• MLflow tracks extraction accuracy using domain expert labels from senior underwriters

Impact

Submission review time drops by 70%. Underwriter capacity doubles because they focus on judgment calls, not data extraction. Risk scoring becomes consistent across all underwriters because the same models and features apply to every submission.

The key insight: ai_parse_document eliminates the need to choose a specific external model for multimodal parsing. Document intelligence is built into the platform.