M&A data room and AI: due diligence without exposing the deal

Can you give a data room to an AI?

Technically, yes: a generative AI can read hundreds of contracts, flag change-of-control clauses, rebuild the cap table history and draft the question list for the seller far faster than a team of associates. The problem is not the reading, it's the sending: uploading the data room as is to an online service means handing a third party the identity of the target, the parties, the employees and the deal terms.

The practical rule: the AI gets a copy in which nobody can recognize the target, the parties or the people, and the original stays with you. To flag that a distribution agreement terminates on a change of control, the AI doesn't need the target's name; it needs the same counterparty to carry the same placeholder in all forty contracts where it appears. Hence batch processing of the entire folder rather than file by file.

Which rules apply to a data room shared with AI?

Four sets of rules overlap. None of them bans AI as such; all of them limit what you may send it.

For lawyers and chartered accountants, professional secrecy comes on top: see generative AI and professional secrecy.

What does the deal's NDA allow?

Every NDA is different, so read yours, together with the virtual data room's terms of use, which often add their own bans (downloading, printing, copying). Five clauses decide what you can do with an AI.

• The definition of confidential information. It often covers not only the documents but also the analyses, notes and summaries you derive from them. A pseudonymized copy is then still derived information, covered by the agreement.
• Permitted recipients. Most NDAs limit disclosure to listed “representatives”: officers, employees, advisers, financing sources. If an AI provider isn't among them, sending documents in the clear is not covered.
• Purpose. Use is usually limited to evaluating the transaction: a service that trains its models on your content falls outside it. Pick a plan that rules this out by contract.
• Return and destruction. When talks end, you will often have to destroy everything, including conversations and files kept by the AI provider.
• Clauses on digital tools. If your NDA mentions AI or online services, it prevails.

If in doubt, ask the seller or the sell-side bank for written consent and describe your method: “only a pseudonymized copy, made and reviewed on our machines, with no company or personal names, is submitted to an AI tool covered by a data processing agreement.” That is far easier to obtain than permission to upload the data room as is. To draft your own agreements with AI: drafting an NDA with AI from your precedents.

MAR: when is the project inside information?

The EU Market Abuse Regulation (MAR) applies to financial instruments admitted to or traded on a regulated market, a multilateral trading facility (MTF) or an organized trading facility (OTF) (Article 2). A deal between two private companies falls outside it; as soon as the target, the buyer or the seller has securities traded on those venues, the question arises.

Article 7(1)(a) defines inside information as “information of a precise nature, which has not been made public, relating, directly or indirectly, to one or more issuers or to one or more financial instruments, and which, if it were made public, would be likely to have a significant effect on the prices of those financial instruments or on the price of related derivative financial instruments”.

• Protracted processes. Intermediate steps (negotiations, a letter of intent, data room access) can themselves be inside information if they meet the criteria on their own (Article 7(2) and (3)).
• The reasonable investor test. Information is price-sensitive if “a reasonable investor would be likely to use” it as part of the basis of investment decisions (Article 7(4)).
• Unlawful disclosure. It occurs when a person holding inside information discloses it to another person, “except where the disclosure is made in the normal exercise of an employment, a profession or duties” (Article 10). Article 14 prohibits it.

Issuers also keep insider lists that include advisers and accountants with access to the information (Article 18). And since June 5, 2026, Regulation (EU) 2024/2809 (the “Listing Act”) has relieved issuers from publishing the intermediate steps of a protracted process, while requiring them to keep that information confidential until it is disclosed (MAR Article 17(1) and (1a)).

What should you mask in a data room?

The target's name is only one of the threads leading to the deal. Here is what a typical data room contains, and what automatic detection covers.

The risk comes from combinations rather than single items: “aerospace parts maker, 140 staff, two plants in southern France” may be enough. See re-identification risk and how to measure it.

How do you pseudonymize an entire data room, step by step?

1. Get sign-off. NDA, data room terms, compliance if a listed issuer is involved, your firm's AI policy. Record who approved what.
2. Export the data room locally. Keep the index structure (1. Corporate, 2. Contracts, 3. HR…). Unzip archives: a .zip is not processed; it is excluded and flagged.
3. Prepare your term lists. “Always mask”: brands, products, domain names, the old code name, site names. “Never mask”: public authorities whose mention reveals nothing. Lists can be imported as .csv.
4. Process the whole folder in one go. AnonymFiles walks the folder and all its subfolders: PDFs (scans included, read by on-device OCR), Word, spreadsheets, presentations, .eml and .msg emails with their attachments. The same company gets the same placeholder in every document, whatever the spelling.
5. Assign code names. In the view that lists the folder's people and organizations, give the target a custom role (CIBLE), and likewise the seller and the buyer: [ORGANISATION_3] becomes [CIBLE_1] in every document. You can also rename a placeholder, for instance [CIBLE_ATLAS]. A name that reuses masked information is refused.
6. Review side by side. Original and copy, document by document, with a re-identification risk indicator (low, medium, high). Mask a missed item, an amount or a passage with one click; draw an area over a logo or signature. Nothing is exported without your approval.
7. Export. The folder structure is kept, files get neutral names (piece-0001…), an index lists exported and excluded documents, and an instruction file for the AI (AGENTS.md) explains the placeholders. Before publishing, every file is reread for the original values: a single hit cancels the export.
8. Hand the copy to the AI. Use a business plan covered by a data processing agreement, with no training on your data. AnonymFiles calls no AI and indexes nothing: your AI tool reads the copy.
9. Re-inject the real names on your computer. The due diligence memo written with [CIBLE_1] gets its real names back in the app, as text, Word or PDF. The encrypted mapping table never leaves the machine: see the pseudonymization mapping table.

The placeholders themselves use French category names, such as [PERSONNE_1] (person) or [ORGANISATION_1]; the AGENTS.md instruction file explains them to the AI.

Which prompts work for due diligence on a pseudonymized copy?

Four habits make answers checkable: explain the placeholders, forbid guessing identities, require the document reference (piece-0001) and an exact quote, and split the work by topic.

Inventory and missing documents

Here is the pseudonymized export of a data room. Bracketed placeholders such as [CIBLE_1] or [PERSONNE_4] replace real names: keep them exactly as they are and do not try to guess who they refer to. Build an inventory by topic (corporate, commercial contracts, financing, employment, real estate, intellectual property, litigation, compliance). For each document: reference, type, date, parties (placeholders). End with the documents usually expected in a due diligence that are missing here.

Change of control and sensitive clauses

Across all contracts in the data room, list the change-of-control, exclusivity, non-compete, non-assignment and early termination clauses. Present a table: document reference, counterparty (placeholder), clause quoted verbatim, consequence if the shareholding of [CIBLE_1] changes, risk level (low, medium, high). Do not invent any clause: if a document is illegible or incomplete, say so.

Cap table history

From the articles of association, shareholder meeting minutes and share transfer forms, rebuild the shareholding of [CIBLE_1] in chronological order. Flag any inconsistency between documents (number of shares, dates, holders) with exact references.

Questions for the seller

Based on your previous analyses, draft the list of questions for the seller, grouped by topic and priority, stating for each question the document that prompts it.

What should you check, before sending and in the answers?

Before sending, review first:

• scanned documents flagged as uncertain reads (rotated page, handwriting, few words recognized);
• logos, stamps and signatures, in letterheads and in schedules;
• spreadsheets: hidden sheets and rows in the original become visible in the copy, so review them too;
• brands, product names and site names that aren't company names;
• combinations of figures, places and dates flagged by the risk indicator.

In the answers, check every quote against the referenced document, and every amount and date. Ask how many documents the AI actually read: a data room often exceeds what a tool handles in one conversation, and an AI that hasn't seen a contract doesn't always say so.

What are the limits of the method?

• Pseudonymizing is not anonymizing. The copy can still be linked back to people through the mapping table: under the GDPR it remains personal data (Recital 26). See anonymization vs. pseudonymization.
• Neither detection nor amounts are 100% automatic. Review remains mandatory, and amounts, which aren't detected, must be weighed document by document.
• File formats have limits. Archives are excluded; 50 MB per document (10 MB for spreadsheets and text); 2,000 pages per PDF, of which 500 can be read by OCR; OCR in French and English. Images in Word files and charts in spreadsheets are not copied: an org chart drawn as an image disappears from the copy.
• The law doesn't change. A pseudonymized copy is no way around the NDA, nor a license to take inside information outside its proper circle.

For the customer due diligence files that often come with an acquisition, see reviewing KYC and AML files with AI. To choose an AI plan: Claude, ChatGPT, Mistral or Copilot for a client file.