Use cases ·

AI for data room review: how accurate is it?

AI for data room review answers due diligence questions with a link to the full document and page, and accuracy varies widely between systems. On FinanceBench, a public test of 150 analyst questions on company filings, Hippocampus got 92% right and OpenAI File Search 52%.*

On this pageWhat is AI data room review?
  1. What is AI data room review?
  2. How does AI data room review work?
  3. How accurate is AI on financial and legal documents?
  4. Can a deal team rely on AI answers in due diligence?
  5. Can AI agents do the first pass of due diligence?
  6. Is it safe to put a data room into an AI system?
  7. Where does AI data room review not help?
  8. How Hippocampus handles data room review
  9. Try it on your own data room

What is AI data room review?

AI data room review is the use of AI to answer a deal team's questions from the documents a seller has shared in a data room. Most diligence questions are look-ups: a figure in one report, a clause in one contract. The slow part is finding the passage and tying it to a page before anyone relies on it. AI does the finding, and the team does the reading and the judging.

How does AI data room review work?

AI data room review works in five steps.

  1. You import the data room, scanned pages included.
  2. The system reads each document once and builds a memory.
  3. The team asks its diligence questions in plain language.
  4. Each answer links to the full document and the page.
  5. If the documents hold no answer, the system says so, which gives you a question for the seller.

AI accuracy on financial and legal documents depends on the system, and the gap is large. On FinanceBench, 150 analyst questions on 84 US company filings, Hippocampus got 92.2% right. Azure AI Search got 76.6%, OpenAI File Search 52.3% and Vertex AI Search 40.6%.* On LegalBench-RAG, 776 questions on 72 contracts and policies, Hippocampus got 85.7% and the next best system 78.3%. The same model wrote the answers in each run, so the difference comes from which passages each system found.

Can a deal team rely on AI answers in due diligence?

A deal team can rely on an AI answer in due diligence when it can check that answer against the source. The link to the document and page is what makes that possible: the associate opens the page, reads the clause and signs off. An answer without a source has to be verified from scratch, which saves no time. A refusal is useful too, because it gives the team a gap to raise with the seller.

Can AI agents do the first pass of due diligence?

AI agents can do the first pass of due diligence when they have a document memory to ask. An agent works through the diligence request list question by question, calls the memory through an API and collects each answer with its document and page. The questions the documents do not cover become the list of gaps for the seller. The deal team gets a first draft of the findings, checks the links on the points that matter and makes the judgement calls.

Is it safe to put a data room into an AI system?

Putting a data room into an AI system is safe if the documents stay under your control and your NDA allows it. Ask the supplier where the documents are stored, whether they are used to train models, and who can read them. Each deal belongs in its own workspace with its own access, so one deal team does not see another's documents.

Where does AI data room review not help?

AI data room review does not judge what a finding means for the deal. It finds the change-of-control clause, and the lawyer decides whether it is a problem. It cannot answer from documents the seller has not shared, and it does not replace management interviews. Its value is that the look-ups get done early, so the team spends its hours on judgement.

How Hippocampus handles data room review

Hippocampus reads the data room once and builds a memory the deal team asks. The team asks in the web browser, or from Claude, ChatGPT and other assistants through an MCP server. Agents reach the same memory through the API. Each answer links to the full document and the page, and Hippocampus refuses a question the documents do not cover. Each workspace has its own search collections and encryption keys, and you run it EU-hosted, in your own cloud or on your own servers.

Try it on your own data room

Create a free account, import your documents and ask your diligence questions today. The Free plan holds 2,000 pages and needs no card. The Starter plan holds 25,000 pages for EUR 30 a month. Larger teams get the Corporate plan, with custom pricing and the option to host it themselves.

Get started free or book a demo.

* GPT-5.4 mini answering. See Evidence.

By Andreas Dahlström, Head of Deeptech Team, Co-Founder

  • 2026
  • Use cases