Choosing pseudonymization software as a French lawyer (2026)
There is no single best tool. The tools French lawyers use to pseudonymise a file before generative AI fall into three architectures: software installed on the computer, browser extensions, and online platforms. The right one depends first on where you accept the original document being processed, then on formats, consistency across a file and evidence of accuracy. Here is what each architecture implies, and the questions to ask any vendor.
Transparency: this page is published by Hexagone AI, a desktop application, so in the first family. We describe architectures, not other vendors' products: check each vendor's current documentation for how its tool works.
Three architectures, side by side
Within each family, products differ. Use this as a map, then check the details with the vendor.
Software installed on the computer
Detection runs on the lawyer's computer. The original document never leaves the firm; there is software to install and keep updated.
Browser extension
The extension sits between you and the AI's web interface. Depending on the product, detection runs in the browser or on the vendor's servers: that is the question to ask.
Online platform
The document is sent to the vendor's servers, often in France or the EU, to be pseudonymised. Nothing to install; the processing has to be framed by contract.
| Question | Software installed on the computer | Browser extension | Online platform |
|---|---|---|---|
| Does the original document leave the firm? | No | Depends on the product | Yes, to the vendor |
| Where is the mapping table? | On the computer | Often in the browser | On the server or sent back to you, depending on the product |
| Processing agreement (GDPR art. 28)? | Not for the pseudonymisation itself | Yes if detection runs on a server | Yes |
| Installation | Software to install and update | An extension | None |
| Typical workflow | Whole files, large documents, any AI tool | Text and attachments sent through the AI's web interface | Documents uploaded one by one or in batches |
Seven questions to ask any vendor
Where is the original document processed?
On your computer, in your browser, or on the vendor's server. The CNB guide (2026) names, as the only ways to remove sovereignty risks, "prior anonymisation or pseudonymisation of the data" or installing the model on the firm's own systems; the Paris Bar's white paper mentions "a locally deployed tool, with no external transmission of data to third-party servers". If the tool is online, the pseudonymisation itself is outsourced processing that needs a contract.
Where does the mapping table live?
It is the key that allows re-identification. The GDPR requires it to be kept separately and protected (article 4(5)). On the lawyer's computer it exists only in the firm; server-side, it depends on the vendor's security.
Formats and scanned documents
A litigation file is scanned PDFs, Excel tables and emails. A tool limited to pasted text covers only part of the work.
Consistency across the whole file
If "Martin" becomes "Person A" in one document and "Person C" in the next, the AI can no longer reason across the file. Check that the same pseudonym is kept from one document to the next.
Reversing the answer
The note the AI produces needs the real names back to be usable. Depending on the tool, that is automatic, one click, or manual.
Measured accuracy, on a named test set
The CNB warns about indirect identification by cross-referencing data. A detection rate means nothing without the test set it was measured on; a public corpus lets you check it. Also test the tool on a real closed file.
Price and billing model
Per user, per page, per word or per office: the real cost depends on your volume. A per-user subscription is predictable; per-page pricing can be cheaper for occasional use.
Sources: CNB, "La déontologie et l'intelligence artificielle" (2026), Paris Bar, White paper on AI (2025).
When a local tool is not the best fit
- You cannot or do not want to install anything on the computer: an extension or an online platform will be simpler to roll out.
- You only work inside ChatGPT's or Claude's web interface: a browser extension follows that workflow more closely.
- Your computer is old or underpowered: local software needs memory and processing power. Hexagone AI, for instance, only supports Apple Silicon Macs (M1 or later) and 64-bit Windows 10/11.
- You process a few pages a month: a free or per-page plan can cost less than a per-user subscription.
Accuracy figures published by different vendors are measured on different test sets and cannot be ranked against each other. Ours are on the Research page, each with its dataset and method.
Frequently asked questions
- What is the best pseudonymization software for a French lawyer?
- There is no single one: the right choice depends first on where you accept the original document being processed. If nothing may leave the firm, look at software installed on the computer. If you only work inside the AI's web interface, a browser extension, after checking where detection runs. If you accept outsourced processing hosted in France or the EU, an online platform. Then test two tools on the same closed file.
- Is a local tool safer than a cloud tool hosted in France?
- It removes a step: with a local tool, neither the original document nor the mapping table is sent to anyone, so there is no processor to assess for the pseudonymisation. A cloud tool hosted in France can be entirely serious, but its security rests on the contract, the actual location of processing and the announced retention, which the CNB guide asks lawyers to check.
- Can vendors' announced detection rates be compared?
- Not directly. Each vendor that publishes a figure measures it on its own test set, with its own rules (some leave dates unmasked, for instance). A figure only means something if the test set is named, ideally public. The most telling test is still to run the same closed file through two tools and review the output.
- Is pseudonymised data still personal data?
- Yes for the firm, which holds the mapping table: the GDPR still applies. The CJEU held in September 2025 (case C-413/23 P) that, for a recipient with no means to re-identify, pseudonymised data may not be personal data. Pseudonymisation remains above all a security measure cited in GDPR articles 25 and 32.
- Is a data processing agreement needed with the pseudonymisation tool?
- If the tool receives your documents on its servers, yes: the vendor processes personal data on your behalf (GDPR article 28). If the software runs entirely on your computer and transmits no document, like Hexagone AI's desktop app, there is no transfer to frame for the pseudonymisation itself.
Try it on your own file.
One week free, no credit card. Nothing is uploaded: run a closed file through Hexagone AI and review the result.
macOS Apple Silicon and Windows 10/11