Professional confidentiality · GDPR compliance
At last, the right to use
frontier AI models*
on your real client files.
Until now, confidentiality rules pushed you towards a cut-down assistant built "for your profession". ndaact.ai swaps names, addresses, IBANs and French social security numbers for placeholders before your text reaches ChatGPT or Claude, then puts the real values back into the reply, on your own computer.
- Keep working with the best models, on your real client files
- Name, date of birth, address, phone number, IBAN: removed from a single prompt in the demo below, and every send is logged
- Your prompts and documents do not go through any ndaact.ai server. The client account is handled separately, see the privacy policy.
For lawyers, notaries, doctors in private practice, DPOs and consultants: professions that answer to a regulator, a patient or a court when client data leaks.
macOS / Windows app + Chrome / Edge extension · free trial · no commitment
* Frontier AI models are the most capable models on the market: ChatGPT and Claude today, with Gemini and Mistral to follow. Until now, confidentiality rules kept them away from real client files.
A client file goes to ChatGPT.
The model only sees placeholders.
Recorded on a Windows PC with a sample client file. The text goes out with placeholders, and the reply comes back with the real values in place. The subtitles explain each step.
Your clients trust you with their secrets.
ChatGPT sends them to OpenAI.
Paste a client file into ChatGPT or Claude and your client's data has left your hands, for a third party you have no control over. For a profession that answers to a regulator, a patient or a court, that is a breach of professional duty first and a technical problem second.
Professional confidentiality
What a client tells you has to stay with you, whichever tool processes it afterwards. If it leaks, the responsibility is yours, whoever built the model.
GDPR: Article 30
Sending a client file to an AI provider is still processing. It needs a lawful basis and, depending on your situation, an entry in your records of processing (see the records guide).
Health data
Once a name sits next to a medical condition, you are handling health data (Art. 9 GDPR). The certified-hosting requirement (HDS, in France) targets hosting done on behalf of the professional; for a consumer AI service the question remains open: see the guide for independent doctors.
Staff use of AI
Your team already uses ChatGPT, with or without your approval. With no record of those prompts, you have a blind spot that you cannot document if you are inspected.
On the third read-through,
one line slips past.
Removing the name, then the ID number, then the IBAN by hand before every prompt takes time, and it pulls you away from the work in front of you.
Worse, it does not hold up. On the third read-through of an exhibit, the night before a hearing, you miss a line and a date of birth goes out. Nobody does this properly fifteen times a day, and that has nothing to do with carelessness.
ndaact.ai does it for you on every send, and the fifteenth time of the day gets the same care as the first.
What happens between your keyboard and the model
ndaact.ai sits between you and the AI model. It redacts your text as it leaves, restores the reply as it comes back, and keeps everything that links the two on your computer.
It finds
Wherever the data sits in your text, it finds the names, the addresses and the identification numbers.
Rules with check-digit validation: NIR (French social security number), IBAN (mod-97), SIRET and bank cards (Luhn), plus email, phone and IP. Then a named-entity recognition model for names and addresses.
It replaces
Each piece of sensitive data becomes a neutral placeholder. The table that links them back is kept on your computer.
Each entity is replaced by a {{TYPE_xxxx}} marker derived with HMAC-SHA256; the mapping table stays local.
It sends
The model receives the markers, not the detected values.
The redacted prompt goes to OpenAI or Anthropic, which receive markers in place of the detected entities.
It restores
The reply comes back readable: the real values are in place again, and you read an ordinary text.
Restoring the original values runs locally when the reply arrives, using the mapping table.
The Chrome and Edge extension,
working with the desktop app
The desktop app covers copy and paste anywhere on your computer. In the browser, text often goes out without passing through the clipboard, so the extension takes over on ChatGPT and Claude at the moment you send.
An add-on to the desktop app
The extension does not work on its own. It talks locally to the app installed on your computer, and the app does the redaction. The extension is included in the Solo plan at no extra cost.
If in doubt, it blocks
If the desktop app does not respond, the send is blocked and you get a warning. There is no fallback mode that would quietly let unprotected text go out.
It works on three websites only
chatgpt.com, chat.openai.com and claude.ai. Everywhere else it reads nothing and does nothing. The rest of your browsing is none of its business.
Available now
Setup takes a few minutes: a desktop app (macOS, Windows) and a Chrome / Edge extension. You are covered for copy and paste, web chat and text-based PDFs, on both your computers. No commitment, and you can cancel whenever you like.
We can hold this price because detection runs on your computer: we pay for no servers to process your prompts, and we do not host your documents.
- ChatGPT and Claude frontier models on the web (Chrome / Edge extension)
- Clipboard (copy and paste in any app) + text-based PDFs (no images)
- 2 devices: your office computer and the one you use in the evening
- Local log retention: 6 months
- Network-wide proxy (all AI tools)
- Several connected computers
- Local log retention, no time limit
- Document pre-redaction (RAG)
- More to come
- Several seats, one invoice
- Payment by bank transfer against a purchase order
- Help with installation and getting started
- A named contact for the rollout
Billed by Inkan.link, in euros. Cancel in one click from your account. See the full pricing page. To get the launch price, ask for a promo code at contact@ndaact.ai.
Where your data goes
What stays with you, what goes to the model provider, and what is stored nowhere.
Stays on your computer
Names, addresses, IBANs, French ID numbers, and the table that links each value to its placeholder.
Goes to the AI model
Placeholders ({{NAME_7f21}}…) in place of the detected entities. An entity that is not detected is not replaced.
Stored nowhere else
Your prompts and documents do not go through any ndaact.ai server. The client account is handled separately, see the privacy policy.
Replacement happens on your computer
Detection and redaction run on your computer. Cloud redaction services, even those hosted in France, first receive your documents unredacted and clean them afterwards. With ndaact.ai, the cleaning happens before sending, on your computer.
Both layers ship together
French regulated identifiers (NIR, SIRET) and IBANs are validated by their check digits. A named-entity recognition model trained on French, bundled with the app, handles names, addresses and company names. There is no second server to install.
Different from an "enterprise" cloud
ChatGPT Enterprise, Azure OpenAI or a self-hosted model shift the risk into the contract. That helps, but the identifying data still leaves your computer. With ndaact.ai, a detected entity is replaced by a placeholder before sending.
A log for Article 30
Each detection is timestamped on your computer with its type and its marker, without the original value. You can show your professional body, your DPO or an inspector what was sent and what was held back.
What the law says
On your side, the mapping table stays on your computer, so the data you hold is still personal data. That reversibility is exactly what makes the reply readable. For OpenAI or Anthropic, which receive only opaque placeholders, the aim is the effect of de-identification.
This follows the relative approach taken by the Court of Justice of the EU in EDPS v SRB (C-413/23 P, 4 September 2025): pseudonymised data may not be personal data for a recipient who has no reasonable means of re-identifying the people concerned.
We do not claim to anonymise the whole prompt. An indirect identifier buried in free text ("my client, the only organic baker in the village") can still point to someone once combined with other information. ndaact.ai redacts only the entities it detects.
Frequently asked questions
What lawyers, notaries, doctors and DPOs ask us before they start using ndaact.ai.
What is ndaact.ai?
ndaact.ai is software that runs on your computer and checks what you send to ChatGPT or Claude. It finds the personal data in your text (names, addresses, IBANs, email addresses, health data, and French identifiers such as the NIR social security number and the SIRET company number), replaces it with placeholders before the text leaves, then puts the real values back into the reply. You use the model you want without giving it your clients' confidential information.
Is my data sent to a server?
Detection and replacement happen on your computer. Your prompts and documents do not go through any ndaact.ai server; the client account is handled separately, see the privacy policy. The text then goes to the model provider, with placeholders in place of the detected entities. Without the mapping table, which stays on your computer, the provider has no reasonable means of finding out who those entities are. An entity that is not detected stays in clear text: detection helps protect your data, without guaranteeing that every entity is found.
Anonymisation, pseudonymisation, de-identification: which one does ndaact.ai do?
The three words mean different things, and the difference matters to a DPO as much as to a lawyer. What decides it is what you actually exchange with the AI.
Technically, ndaact.ai carries out client-side pseudonymisation within the meaning of Article 4(5) of the GDPR: each piece of data it detects becomes a marker, and the table that leads back to the original value stays on your computer. You can reverse the process, and that is what makes the reply readable for you. On your side, the data is therefore still personal data, and calling it "anonymisation" would be inaccurate.
For the model provider and any other third party, the aim is the effect of de-identification: OpenAI or Anthropic receive opaque markers only, without the mapping table, and have no reasonable means of re-identifying the people concerned. This is the relative approach taken by the Court of Justice of the EU in EDPS v SRB (C-413/23 P, 4 September 2025): pseudonymised data may not be personal data for a given recipient who has no reasonable means of re-identification.
In short: pseudonymisation on your side, de-identification as far as the AI is concerned. There is also a limit we prefer to state ourselves. ndaact.ai redacts only the entities it detects, and an indirect identifier buried in free text ("my client, the only organic baker in the village") can still point to someone once combined with other information. So we do not claim to anonymise the whole prompt.
Definitions, legal texts and limits: the guide to anonymisation, pseudonymisation and de-identification.
Is ndaact.ai compatible with professional confidentiality and the GDPR?
ndaact.ai limits what the model provider receives. It replaces detected data with markers on your machine, where the lookup table stays. Each detection is logged locally for your record of processing (Article 30 of the GDPR). Because only detected entities are replaced, read the text before sending it and check what your professional rules allow you to pass to a service provider. ndaact.ai helps with that work without guaranteeing it.
How is this different from a regex filter?
A pattern filter only catches fixed formats: email addresses, IBANs, numbers. It does not read the text, so a person's name, a company name or an address in an ordinary sentence gets through. ndaact.ai combines those rules with a specialised language model that runs locally and picks up those entities as well.
Example: "The file concerning Mrs Sophie Laurent, manager of SCI des Tilleuls, a property company registered at 12 rue de la Paix, Lyon, relates to…". No pattern can flag a person's name, a company name or an address inside an ordinary sentence. A pattern filter, free or paid, sends exactly that to the model in plain text.
Is the AI's reply still readable after redaction?
Yes. The process is reversible: ndaact.ai replaces each piece of data with a placeholder before sending, then rebuilds the reply when it comes back. You get a complete text with the real values in place, while the model receives placeholders in place of the detected entities.
Which professions is ndaact.ai for?
Professions for which confidentiality is a legal duty: lawyers (case files, pleadings, formal notices), notaries (deeds, civil-status records, estate and asset details), doctors in private practice (medical reports, health data), and DPOs or consultants who oversee how their organisation uses generative AI.
Which tools does ndaact.ai work with?
With the assistants in your browser (ChatGPT on chatgpt.com, Claude on claude.ai) through a Chrome and Edge extension, and with everything else on your computer through a desktop app for macOS and Windows. You keep the tools you already use. Redaction happens when you send.
Work on your real client files with the best models
Try it freeAlready a subscriber? Go to your downloads