You drop your submissions into Claude or ChatGPT and a small voice whispers "the head of the bar would not be thrilled". We dug into anonymising documents before sending them to an AI, tested the tools, met the founders. Here is what actually works.
Before panicking: let us put some water in the wine
There is a view doing the rounds at the moment: everything should be anonymised, all the time, or professional disaster follows. It is partly true, but let us set the scene honestly first.
Today, working with American software such as Microsoft or Google already means accepting a compromise on confidentiality. Your mail in Outlook, your files on Drive, your client folders in a shared directory: technically those data pass through American servers, subject to extraterritorial legislation. We live with it, because no European country has managed to build infrastructure as solid and as interconnected as those giants. (And no, I will not talk about Infomaniak. They do very good work, but in terms of day-to-day practicality we are not on the same ground. Hats off all the same to the lawyers who run a whole firm on their tools, that takes nerve.)
So why does AI frighten people so much? Because the sensation is different. Sliding a client file into a chat window feels more intimate, stranger, more exposed than dropping it into a Drive. Except that in practice it is broadly the same logic of data leaving your firm.
Does that mean dropping your guard? No. Remember the summer of 2025: thousands of ChatGPT conversations ended up indexed on Google, some of them full of personal data, because a "make this conversation visible" box was misunderstood. OpenAI pulled the feature urgently and acknowledged there were too many chances to share things accidentally. The incident was fixed. But who can swear it will never happen again?
The reasonable conclusion is therefore neither paranoia nor carelessness. It is a clean layer of anonymisation when you handle documents naming people, and a little common sense on the rest.
The real problem with manual anonymisation
On paper the answer is simple: pseudonymise your documents before sending them to an AI. In real life it is a slog, for three reasons.
Generic tools do not speak law. A run-of-the-mill entity detector does not recognise a company registration number, a case number, a prosecution reference or a bar fund reference. It sees text, not law.
Consistency gets lost along the way. Your Mr Dupont becomes PERSON_001 on page 1, then PERSON_007 on page 12, and the AI ends up tangling itself in its own analysis. Inconsistent anonymisation is sometimes worse than no anonymisation at all.
Re-injection is forgotten. Most home-made fixes leave you with a deliverable full of pseudonyms, which you have to replace one by one with Ctrl+F. Joy.
The ideal workflow, the one the best tools offer, looks like this: you drop in your file (PDF, Word, scans), the tool detects the sensitive entities, you validate by hand what needs validating (a non-negotiable step, no serious lawyer signs without reading it over), you get back a pseudonymised version plus a mapping dictionary kept locally, you work calmly with Claude or ChatGPT on the pseudonymised version, and the tool re-injects the real names into the final deliverable.
The 4 most relevant solutions on the market
We sorted through them. Here are the ones genuinely worth your time.
1. The AI tools built for lawyers (Doctrine, Legora, Harvey, Ordalie, Haiku, Gen-IAL and others)
Most AI platforms built for lawyers already include this building block, or simply do not share your data with the American tools. Often French hosting, new models sometimes integrated within hours of release, an environment built for law. The full package.
The catch for some: when you are used to working inside Claude's interface (with your skills, your MCPs, your projects properly configured) or inside ChatGPT, switching to another tool can break the comfort. That is understandable. But honestly, take the demos. These tools are all genuinely relevant, and you may be surprised.
Who for? Firms that want a complete legal solution, not just an anonymisation layer.
2. Marvin Systems
Probably the most intuitive to date for anyone already working in a Cowork workflow. Marvin sits as an opaque barrier between your documents and the AI assistant: it detects sensitive entities, anonymises or pseudonymises, and restores the real names in one click. It combines in-house named entity recognition, business rules and an open-weight LLM to fit the context of each document. A reassuring point: the files are never stored on their servers, everything is deleted after processing. They are also advised by a lawyer specialising in IP/IT, a member of the Paris bar's AI Law & Ethics Commission.
I met the founders over video. Beyond genuine warmth, the tool is very pleasant and lets you work on your files directly. It does require Claude Cowork, which is a blessing in disguise: for a well-organised firm, it is formidably practical.
- Site: marvin-systems.com/fr
- Pricing: marvin-systems.com/fr/pricing (credit-based, with a trial version you can test)
Who for? Firms already on Claude Cowork who want a smooth anonymisation layer inside their workflow.
3. Hexagone AI
Same spirit as Marvin, but with a strong stance: everything runs locally on your Mac or PC. Detection, anonymisation, mapping, nothing leaves your machine before it reaches Claude, ChatGPT or another model. No copy, no upload, no cloud round trip. They handle procedural documents, opinions, deeds, but also accounting, HR, M&A and health use cases. The position: the only AI layer that respects professional secrecy without forcing you to change tools.
I met them too, the team is excellent and the product holds up.
- Site: hexagone.ai
- Pricing: hexagone.ai/tarifs (free to test, pricing visible)
Who for? The firms most demanding on confidentiality, who want to keep the data physically with them.
4. ReadyForAI
Built by Jimmy Hababou, a lawyer well established in the legal AI ecosystem, who built the tool he needed before making it available to everyone. ReadyForAI pseudonymises, redacts and turns your documents into variables before sending them to ChatGPT, Claude or Gemini, with your data staying under your control. It also lets you build docx templates.
The pricing model is clever: a subscription by number of pages anonymised. That is far more legible than credits, you know exactly what you are paying for. In my view the tool will keep evolving and widening its services.
- Site and pricing: readyforai.fr
Who for? Lawyers who want a transparent price per page and who appreciate a tool built by one of their own.
The real question: do you need all of this?
Here is what we see in our analysis. Plenty of lawyers actually only need the anonymisation block. Not a complete legal AI at 89 or 99€ a month.
The reasoning is simple: they already have their ChatGPT or Claude subscription at 20€, they know it and it suits them. All they are missing is a clean anonymisation layer in front. Paying 100€ a month for an all-in-one when you already have your general-purpose AI can be a lot, especially once you multiply users across the firm.
Hence the real emerging demand: a tool dedicated solely to anonymisation, plugged into the AIs on the market, at a more reachable price. That is exactly the slot Marvin, Hexagone and ReadyForAI aim at, each from its own angle (Cowork integration, local processing, price per page).
The radical path: the local LLM
We mention it for completeness, even though it loses people along the way. Running a model directly on the firm's machine (Ollama with Mistral, for instance) solves the problem at the root: the data never leaves, and anonymisation becomes pointless. The snag: you work with models that are not always up to date, and you give up the power of the latest Claude or GPT. For the vast majority of firms the game is not worth the candle. But the option exists.
What to take away
There is no single answer. The right tool depends on your workflow and on how demanding you are:
- You want a complete legal solution → a tool built for lawyers (Doctrine, Ordalie and so on).
- You are already on Claude Cowork → Marvin Systems.
- You want to keep the data local → Hexagone AI.
- You want a transparent price per page → ReadyForAI.
- You are keeping your ChatGPT or Claude at 20€ as it is and you just want a layer in front → Marvin, Hexagone or ReadyForAI.
And above all, whichever the solution: the manual validation step stays king. The tool detects, you read it over, you sign. The AI assists, the lawyer decides.
We are preparing an even fuller benchmark, with screenshots and field feedback. If you have tested these tools, seen better, or rigged up your own home-made hack, tell us. This is exactly the kind of subject where the concrete detail is missing.
.avif)




.jpg)


