Tools to anonymise data before ChatGPT or Claude: the options compared
If your team wants to use ChatGPT, Claude, Gemini or Copilot on files that contain personal data, there are five broad options. We make one of them, so judge us on the criteria, not the verdict.
- AI enterprise plans improve the contract with the provider, but the provider still receives your data in clear.
- Manual redaction is free and controllable, but slow, error-prone and one-way.
- Presidio is a free, open-source toolkit for developers building their own pipeline.
- Limina (formerly Private AI) is an enterprise de-identification platform deployed by IT, with broad language coverage.
- Hidden Tokens is a desktop app for professionals: consistent codes per case file, values restored in Word and Excel, the key on your computer.
The criteria that matter
- Where the key lives. If the mapping between codes and real values sits with the AI provider or the tool vendor, the protection is weak. It should stay with you.
- Consistency across a case. The same person must get the same code in every document, or the AI will treat one person as several.
- Usefulness of the result. The AI must still be able to reason, and the real values must come back into its output without manual work.
- Proof. You should be able to show a client or a DPO what the AI received.
- Who can run it. A developer toolkit and a desktop app serve different teams.
The options at a glance
| Option | Best for | Key held by | Values restored | Set-up | Price |
|---|---|---|---|---|---|
| AI enterprise plan | Governance of AI use | n/a: data sent in clear | n/a | Admin console | Per seat, from the provider |
| Manual redaction | Occasional short documents | You (by hand) | By hand | None | Free, plus time |
| Presidio | Developers building pipelines | Your system | To build | Engineering project | Open source |
| Limina (formerly Private AI) | Large, multilingual, IT-run deployments | Your infrastructure | In PrivateGPT replies | Containers deployed by IT | On request |
| Hidden Tokens | Professionals using Claude, ChatGPT, Gemini, Copilot on real files | Your computer or file share | Automatic, in Word and Excel | Desktop install, no admin rights | €190 to €390 per seat per year |
Facts about third-party products come from their own websites and documentation, checked on 24 October 2026. If something has changed, tell us at info@hiddentokens.com and we will correct it.
Where Hidden Tokens is not the best choice
- If you need a central service feeding many internal systems, a platform such as Limina, or a pipeline built on Presidio, fits better.
- If most of your documents are in languages other than French or English, check coverage carefully: our engine was built first for French and Moroccan documents, and our benchmark shows it still over-codes ordinary English words.
- If you only redact a few short documents a month, a careful manual method may be all you need.
Where it is
When people who are not developers need to use the AI tools they already have on real client, patient or employee files, with codes that stay consistent across a whole case, the real values restored in Word and Excel, the key kept on their own computers and a record of what the AI saw.
Frequently asked questions
What should I ask any tool before using it with client data?
Four questions: where does the key that links codes to real values live; are codes consistent across a whole case file; how are indirect identifiers handled; and can you prove what the AI actually received?
Is there a tool that guarantees nothing ever leaks?
No. Every automated tool can miss something. Look for one that shows you a plan before coding, lets you correct it, and gives you a way to check what was sent.