FAB Financial Due Diligence Benchmark: Can AI Agents Handle Financial Due Diligence?

Can AI Take Over the Grueling Work of Financial Due Diligence?

Behind the gears of the global economy lies a mountain of complex, painstaking financial work. Financial Due Diligence (FDD) is one of the most demanding disciplines in this space — analysts must extract critical facts from towering stacks of financial statements, contracts, and data room documents, then synthesize them into a complete and reliable assessment. It's long been considered a domain that depends heavily on human expertise.

So just how far can today's AI Agents take this? A team of former Ernst & Young (EY) employees has an answer: they built FAB (Finance Agents Benchmark), specifically designed to test AI Agent performance within real-world financial due diligence workflows.

What Does FAB Test — and What Did It Find?

FAB's core goal is to deconstruct the actual work behind financial due diligence into evaluable tasks, then put AI Agents through their paces. This isn't simple Q&A — it simulates the multi-step analysis of a real consulting engagement: the Agent must locate relevant materials within a data room, extract facts, and then complete analytical reasoning based on those facts.

The team's key finding is telling: Agents can locate relevant facts, but still struggle to weave those facts into a complete, reliable analysis. In other words, current AI is already quite capable at the "information retrieval" layer — but at the "synthesized judgment and end-to-end delivery" layer, it falls meaningfully short of a competent due diligence analyst.

This finding aligns with broader industry observations about large model Agent capabilities: models excel at discrete, atomic tasks, but as the task chain lengthens and requires cross-document integration with logical consistency, reliability degrades. Financial due diligence is precisely this kind of long-chain, high-consistency-requirement scenario — making it an ideal proving ground for measuring an Agent's true capabilities.

Data Room is a central concept in financial due diligence, referring to a controlled collection of documents that buyers or third-party advisors access during a transaction. It typically contains the target company's financial statements, audit reports, key contracts, legal documents, management presentations, and more. Traditional data rooms were physical spaces; modern transactions use Virtual Data Rooms (VDRs) to organize and share these sensitive files electronically. For AI Agents, the data room simulates a real information retrieval environment: large numbers of documents, diverse formats (PDF, Excel, scanned files, etc.), with critical information often scattered across multiple documents in different locations — requiring cross-document comparison to reach credible conclusions. This is the source of FAB's difficulty — Agents must not only understand individual files, but locate, integrate, and verify information across an entire data room.

Why a Former EY Team?

The team's background is worth noting — they are former Ernst & Young employees. Big Four accounting firm practitioners know better than anyone what a real due diligence workflow looks like: which documents matter, how analysis should be structured, and what cross-verification is needed before a conclusion can be considered reliable.

This domain expertise is critical to benchmark quality. The most common mistake in AI evaluation for finance is using overly simplified or impractical tasks to measure models, generating impressive-looking scores that have nothing to do with real-world business needs. Having genuine practitioners design the tasks and data rooms makes FAB far more aligned with actual delivery standards — and its evaluation results far more meaningful.

Ernst & Young (EY) is one of the global "Big Four" accounting firms (alongside Deloitte, PwC, and KPMG). The Big Four conduct a large volume of financial due diligence work in major transactions worldwide — including M&A, IPOs, and private equity investments — with highly standardized, battle-tested methodologies. A typical FDD report covers modules such as Quality of Earnings analysis, working capital normalization, liability identification, and accounting policy assessment, with every conclusion traceable back to specific source documents or policies. Having been deeply involved in hundreds or even thousands of due diligence engagements, Big Four practitioners have a far more intuitive sense than academic researchers of "what AI output is usable versus what is dangerous" — making evaluation criteria they design much more aligned with real delivery quality thresholds, rather than simple information extraction accuracy.

Open, Reproducible — But Not Cheap

FAB takes a fully open approach, with both code and data released publicly:

• GitHub Repository: https://github.com/SecondState-ai/finance-agents-benchmark
• Data Room & Task Set (Hugging Face): https://huggingface.co/datasets/secondstate/finance-agents-benchmark-traces

Open-sourcing means any researcher or developer can reproduce the tests, verify findings, or plug in their own Agent for evaluation. For a benchmark in such a specialized vertical domain, this level of transparency is rare.

The team also candidly acknowledges that building and running this kind of benchmark isn't cheap. They've therefore chosen a phased expansion approach: gradually covering more companies and testing more models. This pragmatic path reflects both the real cost of high-quality financial evaluation (data preparation, task annotation, and multiple model runs all require significant investment) and the fact that FAB is an actively evolving project — not a one-time deliverable.

Implications for AI Agent Adoption in Finance

FAB's value lies not just in producing a score, but in clearly drawing the boundary of current capabilities. For teams looking to bring AI into financial due diligence, auditing, or investment research, here are some practical takeaways:

• Retrieval and information location can confidently be delegated to Agent assistance, with significant efficiency gains;
• Synthesized analysis and final conclusions still require human expert oversight — end-to-end outputs should not be trusted blindly;
• When evaluating AI capabilities in specialized domains, use benchmarks grounded in real business workflows — not generic leaderboards.

As FAB expands to cover more companies and models, it has the potential to become a meaningful measuring stick for progress in financial Agents. And the current picture it reveals — "finds the facts, can't do the analysis" — also points the way for the next stage of model training and Agent design: how to keep Agents logically coherent and reliably conclusive across long-chain tasks will be the key to cracking professional finance use cases.