Pseudonymization software for lawyers: how to choose in 2026

Written by the publisher of AnonymFiles. For each competitor, we only use what the vendor publishes on its own site, as recorded in our comparison of anonymization software on October 14, 2026, and linked in the sources. Prices and features change often: check them before you buy.

Why should a lawyer pseudonymize before using AI?

Because professional secrecy covers the documents in the file, and pasting a brief into an online AI tool hands it to a third party. The AI guide of the Conseil national des barreaux (CNB, the French national bar council, April 2026) recommends never sending information covered by secrecy to a generative AI tool, and “systematically” pseudonymizing or anonymizing data before any use of one. In its section on sovereignty, it goes further: the only ways to eliminate those risks are “prior anonymization or pseudonymization of the data” and installing the model on the firm's own systems (our translations).

Pseudonymizing means replacing each name with a stable label, such as [PERSONNE_1], while keeping on your side the table that puts the names back. That is what lets you get a usable answer back from the AI. The full method is in pseudonymize a document for AI, and the ethics framework in ChatGPT and lawyers' professional secrecy.

What are the three architectures?

Hybrids exist: installed software that also offers a web app, an API or a layer displayed on top of the AI tool. In that case, ask each architecture's questions for every mode you will actually use.

Where is the file processed?

This comes first, because it shapes every other answer. An online platform receives the document before pseudonymizing it: even hosted in France and encrypted, it has had the documents in clear text. For a file covered by secrecy, that is one more third party to govern. Installed software reads the document on the computer, and only the copy you choose to send leaves it.

But “local” does not tell the whole story. According to their own websites, checked on October 14, 2026, several local tools still talk to a server: license activation or checks, model downloads on first launch, usage counters. None says it includes document content, but for a firm that wants to be able to say “nothing leaves”, the difference matters. Our guide offline anonymization software: what “local” really means details these exchanges, and our offline test procedure lets you check them with a firewall.

Where does the mapping table live?

The table links each label to the real name. The GDPR requires it to be “kept separately” from the copy and protected (Article 4(5)), and the EDPB treats it as personal data in its own right. Three situations:

• On the computer, encrypted: the easiest to defend. Check the algorithm and where the key is stored; the CNB guide expects encryption at rest “such as AES-256”, with the key held and managed by the firm.
• In the browser: it can be wiped along with browser data, or synced to a personal account if sync is on. Ask how it is protected.
• On the vendor's server: the vendor then holds both the copy and the means to reverse it. Pseudonymization protects you from third parties, not from the vendor.

Rules on storage, access, retention and deletion are covered in pseudonymization mapping table: the GDPR rules in practice.

When do you need a data processing agreement (Article 28)?

Article 28 of the GDPR requires the controller, here the firm, to use only processors “providing sufficient guarantees” and to bind them by contract. Under the EDPB's Guidelines 07/2020, two basic conditions make an entity a processor: being separate from the controller, and processing personal data on its behalf (paragraph 76). Hence three cases:

The last point deserves a word. In the SRB ruling of September 4, 2025 (C-413/23 P), the Court of Justice accepted that pseudonymized data may not be personal for a recipient unable to re-identify. But in its Guidelines 02/2026 on anonymisation, adopted for consultation on July 7, 2026, the EDPB states that an entity processing data on a controller's behalf is a processor, and that whether the data is personal is then assessed from that controller's perspective (paragraph 15). Pseudonymization does not exempt you from the agreement with the AI provider; it limits what the provider receives. On that agreement, see ChatGPT and GDPR and our comparison Claude, ChatGPT, Mistral, Copilot: which plan for a client file.

What should you ask a vendor?

Ten questions to ask in writing, and the answer a firm should expect:

1. Where is the original document read and analyzed? On the computer, in the browser or on your servers? What about reading scans (OCR)?
2. What goes over the network, and with what content? License, usage statistics, model downloads, updates. Can all of it be turned off?
3. Where does the mapping table live, how is it encrypted, and who holds the key?
4. Do we need a data processing agreement with you? If so: hosting location, sub-processors, retention period, and the nationality of the host (a CNB criterion).
5. Which formats are actually handled? Scanned PDFs, Word, .msg and .eml emails, spreadsheets, presentations. Does a Word file come out as Word, layout preserved?
6. Does the same person keep the same label in every document of the file? Including “Mr. Dupont”, “J. Dupont” and “DUPONT Jean”.
7. How do we review before export? Are the original and the copy shown side by side? Can we mask a docket number or an amount by hand? Is export blocked until each document is approved?
8. How do real names come back into the AI's answer? As text, Word, PDF? On our computer or on yours?
9. What test set is detection measured on? Which types of information are counted? Vendors' figures cannot be compared with each other.
10. What is the pricing model, and what happens to the table if we stop? Per seat, per page, per document, with or without quotas; and will we still be able to put names back into past exports?

The tool comparison table

The seven document pseudonymization tools from our 2026 comparison, seen through the lens of this guide. Data published by each vendor and checked on October 14, 2026, unless stated otherwise; the score is our eight-criteria grid. Prices are excluding VAT where the vendor says so.

The “Article 28 agreement” column is our reading of the published architecture, not a statement by the vendors. None of the tools in this table is a browser extension; Hexagone AI offers a layer on top of Claude, ChatGPT or Cursor that displays the real values in the answer. The tool-by-tool details, with all eight criteria, are in the full comparison, and the head-to-head with the closest competitor in the Hexagone AI alternative.

When is a local tool not the best choice?

Installed software is not the right answer in every situation, ours included:

• You cannot install anything. A computer locked down by IT, a shared or borrowed machine: an online platform hosted in the EU, with a data processing agreement, may be the realistic compromise.
• You work as you type in an AI's web interface. To pseudonymize a question typed on the fly, an extension or a layer on top of the AI is smoother than a round trip through an export. AnonymFiles has no browser extension.
• Your computer is not compatible. AnonymFiles is designed for Apple silicon Macs and Windows; on Linux, Anoni exists; Tacitum only runs on Windows.
• You want to automate a processing pipeline. To pseudonymize thousands of documents inside an information system, a building block such as Microsoft Presidio deployed on your servers, or an API, fits better than an app built around human review.
• You redact a single PDF now and then. Acrobat Pro's Redact tool is enough, as long as you check the result.
• You are a large organization that needs centralized administration (single sign-on, server deployment): compare the vendors' enterprise offers; at AnonymFiles, more than twenty seats is quoted on request.
• You need it today. AnonymFiles is not out yet; Hexagone AI, Lexanon, Anoni, Tacitum and Occlira are available.

How do you test two tools on a real file?

1. Take a closed, representative matter: a scan, a Word file, a few emails, a spreadsheet.
2. Cut the network during processing, or watch connections with an application firewall, as in our test procedure.
3. Count what was missed at review, document by document: names, organizations, addresses, but also docket numbers, amounts and rare job titles (see re-identification risk).
4. Check the export: the copy-paste test on the PDF, metadata, file names, no table. Our method: check an anonymized PDF in 6 tests.
5. Do the round trip with the AI: do the labels survive, and do the real names come back in a clean Word file?

For the documents a firm typically handles and what to mask by hand, see also anonymize legal briefs before AI, pseudonymization software for lawyers, how AnonymFiles works and pricing.