Research
We build the environments and benchmarks that test AI agents on real financial work: the documents, policies and multi-step decisions behind credit, accounting, operations and insurance. We publish what we learn about where models succeed and where they break.
Insurance Agent Evaluation
- Claims Bench 400 verification tasks across 80 property claim files and seven workflows, evaluated with four decision models.
- Underwriting Is a Workflow, Not an Answer Twelve model configurations evaluated on evidence, rule application, and completed insurance underwriting work.
- Would Your Underwriting Agent Make the Same Decision Twice? Repeated runs reveal whether insurance underwriting agents apply the same rules consistently.