A Government Sponsored Enterprise insurance compliance firm processed 50,000+ policies in one month — and uncovered data errors their manual process never caught.

Within three weeks, Advocate had cleared a compliance backlog that had been growing for a decade — and the workflow that did it now runs continuously. In approximately one month of deployment, they processed more than 50,000 insurance policies — nearly eight times what their team had managed manually across the prior year. In the process, the system surfaced something years of human review had never caught: errors buried in the firm’s own source data.

The challenge

When volume outstrips manual processes

Advocate’s platform supports insurance compliance for large Government Sponsored Enterprise lenders. Like most firms operating at this level of complexity, their team built rigorous manual processes for extracting coverage and pricing data: gathering documents, cross-referencing email threads, and keying values into structured databases.

As volume grew, those processes could not keep up — leaving hundreds of thousands of documents unprocessed and the data inside them untouched.

Building with Felix

Codifying and scaling analyst expertise

This kind of work cannot be handed to an LLM alone. Routing hundreds of thousands of policy documents through a foundation model would be prohibitively expensive, and the outputs would vary with every run — producing no reliable audit trail. Advocate needed their analysts’ expertise codified into software that could run their processes consistently, at any volume.

Working with Felix, Advocate built a workflow around their specific data capture requirements. Felix retrieved policy documents by policy and asset ID, ingested associated email threads and attachments, and extracted key fields — broker and insurer details, addresses, limits, policy attributes — into structured, reviewable outputs. Human checkpoints kept every stage auditable and traceable.

Over time, the workflow developed distinctions that typically require experienced human judgment — differentiating, for example, between blanket premiums covering multiple locations and location-specific premiums tied to individual properties. Each layer of additional logic made the system more precise and the outputs more reliable.

The result

Unlocking value from policy data

Within approximately one month of deployment, Advocate had processed more than 50,000 policies — nearly eight times their prior year’s manual output. Years of documentation that had sat beyond reach were now a complete, structured, analytically useful dataset.