AnonymFilesCompare contract versions with AI without exposing the parties
To compare contract versions with AI without exposing the parties, pseudonymize every version together on your computer: the same company keeps the same tag in each one, so a gap the AI flags is a real change in the text, not a side effect of masking. Get the exhaustive list of changes from Word's own comparison, locally, then ask the AI what those changes mean, clause by clause. Check every gap against the originals.
Why compare contract versions with AI?
Because comparison is slow, mechanical and easy to get wrong. The other side's markup of a 30-page contract may come down to one word in the limitation-of-liability clause, a changed cross-reference in the definitions, a new condition on termination. Word's comparison spots every textual change, but it does not tell you what each one means for your client. That is where AI helps: sorting changes, saying which one shifts a risk, linking a new clause to a definition amended elsewhere.
None of these tasks requires knowing who the parties are. The CNB says so in its April 2026 guide, as a recommendation: a firm using AI to analyze contracts, “before loading the documents into the tool, pseudonymizes or anonymizes the parties (names, addresses, phone numbers)” (p. 20; quotations from French texts are our translations).
When to use it, and with which tool?
| Need | Safest tool | Why |
|---|---|---|
| Exhaustive list of changes between V1 and V2 | Word's comparison, locally | Deterministic: every insertion and deletion becomes a revision mark, nothing leaves the computer |
| Effect of the changes, clause by clause | AI, on pseudonymized copies | It explains, sorts and links changes; it can miss one, hence the Word list |
| Extract a contract's clauses into a table | AI, on a pseudonymized copy | A reading task: purpose, term, termination, liability, governing law |
| Hold the other side's draft against your template | AI, both texts pseudonymized together | The texts do not follow the same outline: Word's word-by-word comparison becomes unreadable |
| Framework agreement, amendments and schedules | AI, the whole file pseudonymized | It spots the amendment that quietly changes a clause, if the tags are shared |
Word's comparison runs from the Review tab: Compare, then Compare Documents, choosing the original and the revised document. Word creates a new copy showing the differences as revision marks, without changing the originals (Microsoft support).
Why pseudonymize all versions together?
So that the comparison is about the contract, not the masking. Processed alone, V1 may make the supplier [ORGANISATION_1], and V2, where a subsidiary appears first, [ORGANISATION_2]. The AI concludes that the party changed: a false gap. Dropped into the same folder, the versions share one mapping table: the supplier is [ORGANISATION_1] everywhere, its signatories keep their tags, and only what really changed shows up as a change. The same logic applies to a framework agreement, its amendments and schedules: one folder, one table.
Then tell the AI which copy is which version. Exported files have neutral names (piece-0001…): say in your prompt “piece-0001 = our draft, piece-0002 = the counterparty's markup”, with no date or original file name that would place the deal.
What should you mask in a contract, and what should you keep?
| Item | Mask or keep | Why |
|---|---|---|
| Parties, signatories, contacts | Mask (detected automatically) | Direct identification; roles (“the provider”, “the customer”) stay |
| Registration numbers, offices, delivery addresses | Mask | A registration number finds the company; numbers checked at review |
| IBANs, emails, phone numbers | Mask (detected by rules) | Direct contact details, useless for the analysis |
| Prices, discounts, caps, penalties | Depends on the task, by hand | The price often identifies the negotiation, but once masked it can no longer be compared. Amounts are not detected automatically |
| Product, site and project names | Mask, one click | They place the deal more surely than the parties' names |
| Terms, time limits, notice periods | Keep | Often where the negotiation happens; rarely identifying alone |
| Signature and effective dates | Mask (private dates detected) | Combined with the sector, they date the deal |
| Clauses, numbering, definitions, cross-references | Keep | The very subject of the comparison |
| Governing law, jurisdiction, statutory references | Keep | Needed for the analysis; a cited code article stays in clear |
For amounts, two options. If the analysis does not need them (rewording a cap clause), mask them: both versions then show a tag, and the AI cannot see how they moved. If it does (checking an indexation formula), keep them, or state the change in relative terms in your prompt (“the cap goes from 100% to 150% of the annual price”). The full checklist is in anonymize a contract before ChatGPT; the CNB guide warns that amounts or internal project names can be enough to identify a client (p. 20).
How to compare two versions of a contract with AI, step by step
- Clean each version. In Word, accept or reject tracked changes and delete comments in each version, or go back to the versions actually sent. You then know exactly which text will be compared.
- Run Word's comparison on the originals. Review, Compare, Compare Documents: the exhaustive list of changes, on your computer. Keep it for the final check.
- Pseudonymize all versions together. In AnonymFiles, drop V1, V2 (and any useful amendments or schedules) into the same folder: parties, signatories, addresses, emails, phone numbers, IBANs, private dates and IDs become tags shared by every version.
- Review side by side. Mask amounts and project names with one click if the task does not need them; rename the parties' tags if useful, for example [ORGANISATION_CLIENTE] and [ORGANISATION_FOURNISSEUR].
- Validate and export for the AI. A Word file gives a redacted .docx with the same layout and a text copy for the AI; comments, tracked changes, images and properties are removed. Paste the instructions provided (“Copier la consigne IA”, copy the AI instruction) at the top of the conversation.
- Ask the AI using the prompts below, saying which copy is which version.
- Cross-check its answer against Word's comparison, then reinject the real names into the final note, on the computer, as text, Word or PDF.
Which prompts extract and compare the clauses?
Paste them after the instructions. [ORGANISATION_CLIENTE] is your client, [ORGANISATION_FOURNISSEUR] the other party: use the tags in your export.
1. The table of differences
piece-0001 is our version, piece-0002 the markup from [ORGANISATION_FOURNISSEUR]. Compare them article by article. For each changed article: number, V1 text and V2 text quoted word for word, type of change (addition, deletion, rewording, new threshold or time limit), which party benefits, and why. Ignore formatting. Do not invent any article absent from both versions.
2. Sensitive clauses first
In these two versions, compare first: limitation of liability and caps, warranties, indemnities, termination and its consequences, term and renewal, confidentiality, intellectual property, non-compete, governing law and jurisdiction. For each, sum up in one sentence what V2 changes for [ORGANISATION_CLIENTE].
3. Knock-on effects
Find changes to definitions or cross-references that alter the scope of other clauses without changing those clauses themselves. For each, quote the amended definition and the articles affected.
4. Clause extraction
Extract from piece-0002 a table: purpose, term and renewal, price and price review (without masked amounts), payment terms, each party's main obligations, liability, termination, confidentiality, governing law, jurisdiction. Give the article number for each item.
5. The negotiation note
From the table of differences, write a one-page note for [ORGANISATION_CLIENTE]: changes to accept, to discuss and to reject, each with a one-line reason. Keep the tags.
For NDAs, see also drafting an NDA from your own precedents, where the same method compares your usual clauses with the counterparty's.
What does the AI give you, and what must you check?
A clause-by-clause table, readable and reasoned, which no set of revision marks provides. But an AI can skip a change, invent one or misread its direction. The CNB guide asks lawyers to “reread, check and validate every generated content” before using or sending it (p. 28).
- Completeness. Every change in Word's comparison must appear in the AI's table, or be set aside by you as purely cosmetic.
- Quotes. The “V1” and “V2” texts quoted by the AI appear word for word in the originals.
- Direction of the change. “Favorable to the client” is a judgment call: yours is the one that counts.
- Masked amounts. The AI never saw them: check how they moved in the originals.
- Tags. After reinjection, each party is in its place: swapping customer and supplier turns the whole note upside down.
Which rules apply?
- French lawyers. A contract negotiated for a client falls under professional secrecy (Act of December 31, 1971, Art. 66-5; national bar rules, RIN, Art. 2.2 for client names). The CNB guide rules out attaching a file that names the client to a prompt (p. 15) and recommends systematic pseudonymization (p. 18). See the CNB guide, point by point.
- In-house counsel. Prices, terms and negotiating positions may qualify as trade secrets, which require “reasonable protective measures” by their holder (French Commercial Code, Art. L. 151-1). Handing the versions in clear text to an online service can weaken that protection.
- Confidentiality clause. The contract itself, or the NDA governing its negotiation, may prohibit disclosure to third parties; and Article 1112-2 of the French Civil Code makes liable anyone who uses or discloses without authorization confidential information obtained during negotiations.
- GDPR. Signatories and contacts are individuals: their data must be “limited to what is necessary” (Art. 5). The AI plan matters too: see Claude, ChatGPT, Mistral or Copilot for a client file.
What AnonymFiles does with a contract and its versions
AnonymFiles, French document anonymization software, is a desktop app coming soon for Apple silicon Macs and Windows. It pseudonymizes a whole folder, Word (.docx), text or scanned PDFs (local OCR), spreadsheets and text files, with the same tag for the same party across every version. Each document is reviewed side by side before export; a Word file comes out redacted with the same layout, and so does a PDF, each masked passage becoming a labeled black box whose text is actually removed. It does not compare versions itself and calls no AI: it prepares safe copies, then puts the real names back into the answer, on the computer. Amounts are masked by hand. 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). See also in-house counsel and DPOs and the use cases.
Frequently asked questions
Can you compare two versions of a contract with ChatGPT without exposing the parties?
Yes: pseudonymize both versions together on your computer, so each party keeps the same tag in both, then give the AI the tagged copies. It compares the clauses without knowing who the parties are, and you put the real names back into its answer on your machine.
Why not pseudonymize each version separately?
Because the same company may then get a different tag in each version. The AI sees a change of party that does not exist. Processed together, the versions share one mapping table, and only what really changed shows up as a change.
Does AI replace Word's comparison?
No, it complements it. Word's comparison, run locally, lists every textual change without sending anything online. AI explains their effect clause by clause, but it can miss one: always cross-check its table against Word's list.
Should prices and liability caps be masked?
If the task does not need them, yes: a price often identifies the negotiation. Once masked, the AI can no longer compare them; state the change in relative terms in your prompt, or check it yourself in the originals. In AnonymFiles, amounts are masked by hand, with one click.
What about tracked changes in a Word file?
Accept or reject the changes in each version before processing it, so you know exactly which text is compared. In the copy exported by AnonymFiles, comments, tracked changes, images and document properties are removed.
Can in-house counsel use AI on a contract under negotiation?
Yes, with a pseudonymized copy. The terms of a contract under negotiation may be trade secrets, protected provided there are “reasonable protective measures” (French Commercial Code, Art. L. 151-1): removing parties, prices and projects before sending is one of them. Check the contract's confidentiality clause too.