Insurance companies must process underwriting cases efficiently, particularly for non-Straight Through Process (non-STP) cases, which often involve manual reviews of medical documents and risk assessments. This traditional process is time-consuming, prone to errors, and challenging to scale. To address these challenges, a genAI-driven underwriting agent using multimodal LLMs can be developed. This system automates the underwriting process, improves decision accuracy, enhances efficiency, and reduces the time required to process complex cases.

The traditional underwriting process for non-STP cases poses several challenges, including:

• Time-consuming manual review of medical documents and proposals
• High risk of delays due to human errors
• Difficulty in scaling due to manual intervention
• Complex decision-making based on various medical conditions and policy requirements

Develop a genAI-driven underwriting engine that automates non-STP underwriting processes. It leverages multimodal LLMs to extract information from medical documents and further utilizes LLMs to assess the risk associated with each proposal. The system categorizes proposals into red, amber, or green based on medical records and raises appropriate queries or suggests appropriate actions.

Key components include:

• Data handling service: Proposal forms and medical documents are uploaded and made available for further processing. 
• Intelligent Document Processing engine
  a. IDP controller:  The primary component responsible for executing all tasks outlined below and generating corresponding results for each task.
  b. Document classifier: Categorizes incoming documents into medical and non-medical classifications to ensure accurate processing and appropriate handling of each document type.
  c. Proposal form processor: Handles complex proposal form formats and its variants to extract text and identify relevant entities, facilitating efficient data processing and integration.
  d. Medical document processor: Specifically designed to handle medical documents, this processor extracts document text and identifies key entities to ensure accurate data capture and analysis of medical content.
  e. Non-medical document processor: Specifically designed to handle non-medical documents, this processor extracts document text and identifies key entities to ensure accurate data capture and analysis of medical documents.
• Risk engine
  a. Risk assessment and categorization: The extracted data is evaluated by LLMs, categorizing proposals into red, amber, or green based on predefined risk criteria.
  b. Automated decision making: GenAI-driven analysis recommends approval, rejection, or further specialist review, ensuring accurate and timely decisions.
• Report generation: The system aggregates & generates a report summarizing the medical documents and decisions.

• Azure OpenAI multi-modal LLM  for extracting and processing data from medical documents.
• YOLO (You Only Look Once) for document and table classification.
• LLMs (such as GPT-4 Turbo) for risk assessment and decision-making.
• Python for data processing and workflow automation.
• MongoDB for intermediate and final data store
• Redis for caching
• Celery for task queuing and management
• ReactJS for generating and visualizing underwriting reports.
• Azure Cloud for secure data storage and scalable system deployment.

A genAI-powered underwriting automation system reduces manual effort, speeds up processing, and enhances risk identification. With 24-hour case closure and improved scalability, healthcare insurance companies can streamline their underwriting process, increase straight through processing, improve loss ratios and overall efficiency.