FAB
A financial agent benchmark constructed by a team of former Ernst & Young employees, designed to evaluate AI agent performance in financial due diligence workflows. It covers multi-step tasks including data room document retrieval, fact extraction, and synthesis analysis across complex transactional scenarios.
Timeline (last 90 days)
FAB (Finance Agents Benchmark) was built by a team of former Ernst & Young (EY) employees and is specifically designed to test AI Agent performance in real financial due diligence workflows.
FAB tests multi-step analysis simulating real consulting scenarios, where Agents must locate relevant materials in a data room, extract facts, and complete analytical reasoning based on those facts, rather than simple Q&A.
The FAB team's key finding is that AI Agents are able to locate relevant facts, but still struggle to weave those facts into a complete, reliable analysis.
FAB adopts a fully open strategy, with code open-sourced on GitHub (github.com/SecondState-ai/finance-agents-benchmark) and the data room and task set published on Hugging Face (huggingface.co/datasets/secondstate/finance-agents-benchmark-traces).
Due to the high cost of building and running the benchmark, the FAB team chose to expand in phases, gradually covering more companies and testing more models.
All Facts (5)
FAB (Finance Agents Benchmark) was built by a team of former Ernst & Young (EY) employees and is specifically designed to test AI Agent performance in real financial due diligence workflows.
50%UnverifiedFAB tests multi-step analysis simulating real consulting scenarios, where Agents must locate relevant materials in a data room, extract facts, and complete analytical reasoning based on those facts, rather than simple Q&A.
50%UnverifiedThe FAB team's key finding is that AI Agents are able to locate relevant facts, but still struggle to weave those facts into a complete, reliable analysis.
50%UnverifiedDue to the high cost of building and running the benchmark, the FAB team chose to expand in phases, gradually covering more companies and testing more models.
50%UnverifiedFAB adopts a fully open strategy, with code open-sourced on GitHub (github.com/SecondState-ai/finance-agents-benchmark) and the data room and task set published on Hugging Face (huggingface.co/datasets/secondstate/finance-agents-benchmark-traces).
50%