Fortune 500 insurer cuts quote turnaround 90% with LLM-powered claims data extraction

The insurer relied on traditional manual methods to extract data from unstructured documents such as emails and PDFs, which were time-consuming and error-prone. This led to errors and inconsistencies in claims platforms, data quality issues, and significant inefficiencies in claims processing workflows.

Bitwise implemented an AI solution using Large Language Models (DBRX Instruct and LLAMA3 8B) on Azure Databricks to automatically read and extract key data points from insurance documents. The system uses a three-layer medallion architecture: a Bronze Layer for raw storage, a Silver Layer for AI-powered extraction of coverage amounts, deductibles, and expiry dates, and a Gold Layer for human-validated clean data.

Quote turnaround time dropped 90%, falling from 2 days to 2 hours. The LLAMA3 8B model achieved 85% accuracy in automated data extraction. The solution also delivered enhanced scalability to handle large volumes of unstructured data and increased overall employee productivity.

• A medallion (Bronze/Silver/Gold) architecture with a human-in-the-loop validation step is effective for balancing automation speed with data accuracy in regulated industries.
• Open-source LLMs like LLAMA3 8B can achieve production-viable accuracy (85%) for structured data extraction from insurance documents.
• Automating unstructured document ingestion directly addresses downstream claims platform errors and inefficiencies.