Draft an NDA with AI from your own precedents and templates

Why start from your precedents rather than a generic NDA?

Because a general-purpose model drafts an average NDA, while your library holds what has already been negotiated: a definition of confidential information that held up, carve-outs the other side accepted, a defensible survival period, a calibrated penalty clause, a non-solicitation clause that got through a general counsel's review. Having the AI work from those texts gives you a first draft in your own style and on your own terms.

The catch is that this library is also one of your most sensitive assets. Each agreement names a client, a counterparty, signatories and often the contemplated deal. Many also state that their existence and terms are themselves confidential: sending them in clear text to an online service may breach those undertakings, depending on the wording. And even without an NDA, French law makes liable “anyone who uses or discloses without authorization confidential information obtained during negotiations” (Civil Code, Art. 1112-2; quotations from French texts are our translations).

When should you use AI on an NDA library?

• Building or refreshing a template: start from ten signed agreements and derive a model with options.
• Preparing a first draft for a new deal, from the closest precedents (mutual or one-way, sale, partnership, recruitment).
• Comparing your positions: which carve-outs you accept, which terms you sign, where you gave ground.
• Answering the counterparty's draft: hold its NDA up against your usual clauses, one by one (see also comparing contract versions with AI).

What the AI does not do for you: choose the governing law, read the balance of power, or guarantee that a clause is valid. It suggests; you draft.

What should you keep or mask in an NDA, and why?

Mask what identifies the parties and the deal; keep what makes the agreement work. The CNB's April 2026 guide recommends that, before loading contracts into an AI tool, a firm pseudonymize or anonymize “the parties (names, addresses, phone numbers)” (p. 20), and warns about identification by cross-referencing: “amounts, internal project names, the way a contract is drafted”.

Then there are the Word file's traces: tracked changes, comments (“client refuses the penalty clause”), document properties. In AnonymFiles, a Word file comes out as a redacted .docx with the same layout, with comments, tracked changes, images and properties removed, plus a text copy for the AI. Contract-specific pitfalls are covered in our checklist for anonymizing a contract before ChatGPT.

How to draft an NDA from your precedents, step by step

1. Pick the precedents. Eight to fifteen representative agreements are often enough: the most recent, the most negotiated, from each family (mutual, one-way). Prefer clean signed versions to annotated drafts.
2. Add a deal sheet. A plain .txt file: the parties, their role (disclosing, receiving), the type of deal, the sensitive points. Dropped in with the precedents, it gets the same tags, which is what lets the real names go back into the draft.
3. Pseudonymize the whole folder. Drop it into AnonymFiles: names, companies, addresses, emails, phone numbers, private dates and ID numbers become stable tags. Your client, if it appears in several agreements, keeps the same tag throughout.
4. Review and complete. Side by side with the original, mask amounts, project names and deal descriptions with one click. Rename the useful tags, for example [ORGANISATION_CLIENTE] and [ORGANISATION_PROSPECT]; the name stays in brackets for reinjection.
5. Validate and export for the AI. Once each document is validated, the export gives the tagged copies, and “Copier la consigne IA” (copy the AI instruction) copies the instructions to paste at the top of the conversation. The mapping table stays on the computer.
6. Work with the AI in three passes: clause extraction, template, then draft (prompts below).
7. Reinject and finish in Word. Paste the draft into “Réinjecter un résultat” (re-inject a result): the deal sheet's tags turn back into the real parties, and you save the result as a Word file for the final review. Masked amounts come back as they were; the ones you want to propose are typed in by hand.

Which prompts extract the clauses and build the template?

Paste them after the export's instructions. The tags are examples: use the ones in your file.

Asking for word-for-word quotes and the source agreement of each option makes the work checkable: you find the clause in the original precedent in seconds.

What does the AI give you, and what must you check?

A first draft in your own wording, with a trace of the options chosen, and a table of your negotiating habits. Before anything goes out, review it as you would an associate's draft; the CNB guide asks lawyers to “reread, check and validate every generated content” (p. 28).

Which rules apply to lawyers and in-house counsel?

For French lawyers (avocats)

NDAs drafted for clients fall under professional secrecy, which covers “all the documents in the file” (Act of December 31, 1971, Art. 66-5) and “the names of clients” (national bar rules, RIN, Art. 2.2). The CNB guide rules out attaching to a prompt a file that contains the client's name (p. 15) and recommends systematic pseudonymization or anonymization before any use (p. 18).

Reusing one client's precedents for another also raises a conflict-of-interest question. The guide recalls that a lawyer may not take on a new client if the secrecy of information given by a former client risks being breached (Decree no. 2023-552 of June 30, 2023, Art. 7), and that the lawyer must make sure data sent to the tool “is not at risk of being reused in other matters” (p. 30). A pseudonymized library, stripped of deals and amounts, sharply reduces that risk: what the AI sees is clauses, not clients.

For in-house counsel

The company's agreements and the information they protect may qualify as trade secrets, which require “reasonable protective measures” by their holder (French Commercial Code, Art. L. 151-1): sending the library in clear text to an online service can weaken that protection. Named signatories and contacts are individuals: the GDPR requires their data to be “limited to what is necessary” (Art. 5). Finally, a French Act of February 23, 2026 creates confidentiality for in-house counsel's legal opinions, under conditions and from a date set by decree; it covers opinions, not contracts. See in-house counsel and DPOs.

What are the limits of the method?

• Pseudonymization hides names, not style. A very distinctive clause may stay recognizable to whoever negotiated it. The CNB guide mentions the way a contract is drafted (p. 20): leave out the most unusual agreements.
• Amounts are masked by hand. AnonymFiles does not detect them automatically; you select them at review, once for the whole file.
• Quality depends on the precedents. If your agreements are old or poorly kept, the template will inherit that. Update them before using them as models.
• A draft is not advice. Choice of law, negotiation strategy and the validity of clauses remain your work.

On how reversible pseudonymization works, see pseudonymize a document for AI, then put the real names back and the mapping table.

What AnonymFiles does with an NDA library

AnonymFiles, French document anonymization software, is a desktop app coming soon for Apple silicon Macs and Windows. It processes a whole folder of Word (.docx), text or scanned PDFs (local OCR) and text files, with the same tag for the same company across every agreement. Each document is reviewed side by side before export; a Word file comes out redacted with the same layout. The mapping table is encrypted (AES-256-GCM) and never leaves the computer, and real names are reinjected on the machine. The app calls no AI: you choose the tool and what you give it. Interface in French and English, no account, no telemetry. Pricing: €290 excl. VAT per seat per year, €240 from 3 seats, 14-day trial (pricing). Other situations: use cases.