How to evaluate an AI due diligence vendor: 10 questions

Ask before the data room goes in. Ten questions cover what a buyer needs to know: where the documents are processed, who keeps them, whether the NDA allows it, how each finding is sourced, what the tool did not read, what is missing, which law it knows, who signs, what it claims about accuracy and what drives the price. Ask for every answer in writing. The list works for any vendor, including us.

1. Where are our documents processed?

Data room documents are confidential under the NDA and often hold personal data under the GDPR. Ask which model providers handle them, in which region, and whether that region is the default or an option you have to switch on.

A good answer names the providers and the region, and sits in the data processing agreement.

2. Does anyone keep or train on our documents?

Two separate questions. Does the model provider retain prompts and documents, and for how long? Does the vendor or the provider use them to train a model?

A good answer gives the retention period per provider, says no to training, and puts both in the contract.

3. Does our NDA allow it?

Most M&A NDAs allow disclosure only to named representatives: staff, advisers, financing sources. A software platform and its sub-processors are rarely on that list. Have your lawyer check before the first upload.

A good answer: the vendor signs an NDA before any document moves, and lists its sub-processors.

4. At what level does each finding cite its source?

Document, page, or the passage itself. A finding your reviewer cannot trace has to be redone. Not every finding has a passage to quote: a calculation over a spreadsheet points to the sheet or the cells it used.

A good answer: every finding names its source document and location, with the supporting passage where there is one.

5. Does it read every document, or a sample?

Some tools answer questions by retrieving the passages that look relevant. That is search, and the clause that matters can sit in a document it never opened. Ask what share of the room was read, and what happens to a file it could not read: a scan, a locked PDF, a broken export.

A good answer: every document is read, and anything that could not be read is listed.

6. Does it tell us what is missing?

An absent document leaves no passage to quote, so a tool that only reads what is there will never flag it. The red flag is often the contract nobody uploaded.

A good answer: a separate list of what a buyer would expect and did not find, checked by a person.

7. Which disciplines, and which law?

Financial, commercial and operational reads work across borders. Legal and tax depend on the law of the target. Ask which disciplines are covered in one run, whether they read each other’s findings, and which jurisdictions the legal and tax reads are built for.

A good answer names the jurisdictions, and says plainly where you need local counsel.

8. Who signs off the findings?

AI output is a draft until a person accepts it. Ask whether anything can be exported without a named human sign-off, whether the vendor gives any assurance, and whether a bank or a W&I insurer can rely on the output.

A good answer names the human gate and is clear about reliance. Most platforms give no assurance; that is fine, as long as they say so.

9. What does the vendor claim about accuracy, and who measured it?

A percentage without a method is a marketing number. Ask on which dataset it was measured, by whom, and whether anyone outside the vendor checked it.

A good answer gives the method, the dataset and the auditor, or says plainly that there is no audited rate.

10. What does it cost, and what moves the price?

Per seat, per document, per page or per deal: the model sets what a large or messy room will cost you. Ask what is included, what is extra, and what happens if the room grows during the process.

A good answer is one price for an agreed scope, in writing, before the room goes in.

How does Factum answer the ten?

Briefly, and with the gaps left in.

  1. On EU endpoints by default, under a data processing agreement.
  2. Zero retention at the model provider, which does not train on your documents. Pseudonymisation before analysis. Legal, financial and tax analysis is hard-gated to zero-retention endpoints.
  3. NDA first. Documents move only after it is signed.
  4. Every finding names its source document and location, with the supporting passage where there is one.
  5. Every document in the room is read, across 9 modules in 6 waves.
  6. A list of what never arrived, from the first run on.
  7. 8 disciplines in one run. Legal and tax are tuned to Dutch law; for a target elsewhere, have local counsel check those findings.
  8. Your own reviewer signs off. No export leaves the platform without that sign-off. Factum signs nothing and gives no assurance.
  9. No audited accuracy rate. Factum is pre-launch.
  10. Price on request: one price for the whole scope, set against the size and state of the data room.

Details on how your documents are handled and what AI cannot do in due diligence.

Frequently asked questions

Do these questions apply to a virtual data room with AI features?

Yes. Questions 1 to 3 and 8 to 10 apply to any tool that touches the documents. Questions 4 to 6 matter most for a tool that claims to do the diligence reading itself.

Which question matters most?

Question 3 comes first in time: if the NDA does not allow the upload, the rest does not matter. After that, question 8. Whoever signs carries the findings.

How do we test a vendor’s answers?

Run it on a deal you closed last year. Nothing in process is exposed, and you can hold the findings against what your team found then.

Can AI due diligence replace our advisers?

No. It reads the whole room and shows the specialists where to look. The accountant’s financial due diligence and your lawyer’s legal due diligence stay what they are, and your own reviewer decides what matters.

Where Factum fits

  • For private equity and other acquirers, before the binding offer or the signing.
  • Every document in the room read, with each finding citing its source document and location.
  • A list of what the seller never uploaded.
  • Your reviewer signs off in the platform.

Factum is pre-launch: there is no public client or case study yet. How this differs from the AI in a virtual data room: AI due diligence vs. data room AI. Before signing: red flag review of a data room. Buying a Dutch BV: due diligence on a Dutch BV.

Run it on a deal you have open now, or one you closed

Share the room under NDA, we run the full pass, and you get the findings on a call, source by source. If a sprint is not the right instrument, we will say so.

The Findings Guarantee

Three material findings your team did not already have, or the engagement is free.

For pilot mandates only. The guarantee does not apply to demos.