# Webinar: Building a life new business underwriting workflow with Llamaparse

## Executive summary

This webinar details the architecture of a life underwriting workflow, demonstrating how AI-driven document parsing, splitting, classification, and extraction can automate data ingestion from complex, multi-page medical records. The system emphasizes a Human-in-the-Loop (HITL) approach, utilizing confidence scores and deterministic checks to flag discrepancies, ensuring data reliability before final decision-making.

## Key takeaways

- The Role of the Parsing Pipeline: The workflow relies on a sequence: Parse (converting documents to structured JSON/Markdown), Split (breaking large documents into granular parts), Classify (identifying document type with confidence scores), and Extract (pulling specific key-value pairs).
- Ensuring Data Trustworthiness: The system incorporates critical checks, including checking for missing pages, verifying data consistency across sources, and flagging heavily degraded scans. Confidence scores are provided on a per-attribute basis.
- Model Reliability and Evaluation: To maintain stability in production, models can be 'pinned' after formal evaluation. The company promotes the use of open-source tools like parsbench 2 for running independent evaluations.
- Workflow Automation vs. Human Judgment: AI is used to handle the high-volume, repetitive tasks (reading, checking, cross-referencing), allowing underwriters to focus on the complex, human elements like reaching out to applicants or resolving discrepancies.

## Technical details

- Parsing Tiers and Complexity: Parsing solutions are categorized into tiers: Fast (text-only, digital text layer), Cost-Effective (basic charts/images), Agentic (complex tables), and Agentic Plus (most messy documents, stamps, dense tables). A 'cost optimizer' can automatically drop down the tier if a page is simple enough.
- Parse vs. Extract Distinction: Parse outputs the entire document content (JSON/Markdown) for agent consumption, while Extract builds on the parsed output to pull specific, defined key-value pairs based on a schema.
- Data Storage and Retrieval: Extracted information is stored in both relational databases and vector databases, enabling agents to perform comparisons and allowing humans to query the data via a chatbot with citation links.
- Handling Complex Documents: For large documents, the 'Split' solution breaks the input into granular components (e.g., fax cover sheet, lab results, office notes) before extraction, improving processing accuracy.

## Practical implications

- Automating the review of high-volume, complex documents (e.g., medical records) significantly increases processing capacity (e.g., 5,000 to 40,000 documents).
- The system reduces the need for manual data rekeying by providing structured, searchable data (JSON/Markdown) directly to the user.
- The HITL design ensures that AI findings are always traceable back to the specific word, line, or bounding box in the original document.

## Topics

Document Parsing, AI Workflow, Underwriting, Data Extraction, HITL, LLMs, LlamaIndex, parsbench 2, light parse

Source: https://www.youtube.com/watch?v=kx8ma6pnkBQ
