Develop a Custom AI Underwriting Model

Most SMB insurers and MGAs run on an AMS like Applied Epic, Vertafore, or HawkSoft. These platforms are excellent systems of record for policy administration but offer rigid, limited capabilities for custom underwriting logic. You cannot easily add a new proprietary risk factor for a niche market or change the weighting of existing factors without a costly and slow vendor engagement.

To compensate, senior underwriters build complex risk models in Excel. A 15-person MGA specializing in commercial property might receive 50 submissions a day, each as a bundle of ACORD forms, supplemental PDFs, and loss run reports. A junior underwriter spends 30 minutes per submission manually transferring data from these documents into the AMS and the master Excel rating sheet. This process is slow, creates a bottleneck for quotes, and is highly susceptible to data entry errors that impact pricing accuracy.

The structural problem is that an AMS is built for data storage, not computation. Excel is a calculation tool, not an auditable, multi-user production system that can process unstructured data. Off-the-shelf insurtech tools are often designed for large carriers with standard workflows and high monthly minimums. They provide a black-box model that you cannot inspect or tune to the specific risk profile of your book of business, and they rarely integrate cleanly with an older AMS.

The first step is a discovery process focused on your data and underwriting guidelines. We would start by auditing your last 12 months of submission data (applications, supplemental forms, loss runs) and your existing rating spreadsheets. This audit maps every data point to its role in the risk assessment, identifying the most predictive signals and confirming data quality. You would receive a clear data plan before any code is written.

The core of the system would be a Python service built with FastAPI and deployed on AWS Lambda for efficiency. When a new submission package arrives via email or an AMS webhook, the Claude 3 Sonnet API parses all documents, including PDFs and images, to extract structured data. Pydantic models validate this information before it is fed to a lightweight risk model. This serverless architecture ensures processing costs remain low, often under $50 per month, and scales automatically with submission volume.

The delivered system exposes a secure API endpoint that returns a risk score from 0-100, a recommendation (accept, decline, refer to senior underwriter), and an explanation of the top 5 factors that influenced the score. This output can be written back to a custom field in your AMS. You receive the complete source code in your GitHub repository, a runbook for model retraining every 6 months, and full ownership of the system.