Improve Risk Assessment with a Custom AI Underwriting Model

Independent agencies run on their Agency Management System (AMS), whether it is Applied Epic, Vertafore, or HawkSoft. These platforms are excellent systems of record for structured ACORD data, but they treat critical supplemental documents like loss runs, financial statements, or contractor questionnaires as opaque files. The most important risk indicators are trapped inside these PDFs, invisible to the AMS.

Consider an underwriter at a 15-person agency reviewing a general contractor's application. The ACORD 125 form is clean, but a supplemental PDF mentions previous work involving 'crane operations' and 'subsurface excavation'. An experienced underwriter knows these are major red flags. They must then spend 30 minutes manually searching for the contractor's OSHA violation history and reading prior loss runs for related claims. A less experienced or rushed underwriter might miss these textual cues entirely, leading to a mispriced policy.

This manual process is the failure mode. Your AMS has no mechanism to read the text in an attached PDF and flag keywords. You cannot create a rule in Applied Epic that says, 'If the word 'crane' appears in any document, increase the preliminary risk score by 20 points.' The system's logic is confined to the structured data fields, forcing your most expensive talent to perform low-level document review on every single application.

The structural problem is that an AMS is built for record-keeping, not predictive analysis. Its data model is fixed by the vendor. You cannot add new data fields to track emerging risks specific to your niche, like cybersecurity protocols or supply chain dependencies. This forces agencies into a reactive posture, relying on the institutional knowledge of individual underwriters, which is inconsistent and does not scale.

The first step is a data audit. Syntora would connect to a sandboxed copy of your AMS to analyze the last 24 months of applications, both bound and declined. We would work directly with your underwriters to map the unstructured data they review manually (like supplemental narratives) to policy outcomes. This discovery process produces a report identifying the top 15-20 predictive features hidden in your documents and a clear plan for the model build.

We would build the technical system using a series of AWS Lambda functions to form a data processing pipeline. When a new document is attached to a policy in your AMS, a webhook triggers the first function. The Claude API parses the document text, extracting key entities and clauses. This structured output is then fed into a Python-based scoring model that calculates a risk score. The final result is written back to a custom field in your AMS via a FastAPI endpoint.

The delivered system provides an immediate, actionable signal within your existing workflow. Your underwriters would see a 0-100 risk score and a one-sentence summary (e.g., 'High risk: Prior water damage claims and mention of subcontractor work') directly on the application screen in Vertafore or Applied Epic. This allows them to triage their work effectively, spending their valuable time on the 20% of applications that contain 80% of the risk.