Claude, ChatGPT, Mistral or Copilot on a client file: what leaves, what stays

The summary table: what does each plan do with your data?

Each cell links to the vendor’s official page that supports it, checked on October 19, 2026. Where a vendor publishes nothing on a point, the cell says “not publicly documented”. These terms change often: re-read the pages cited before you decide.

How to read the columns. Training means reusing your conversations to improve the vendor’s models. Retention is how long the vendor keeps your data, including after deletion. The DPA (data processing agreement) is the processor contract GDPR Article 28 requires when a provider handles personal data on your behalf. Zero data retention (ZDR) means the vendor doesn’t store prompts and outputs beyond processing.

How should you read the table, vendor by vendor?

ChatGPT (OpenAI)

The dividing line runs between individual and business plans. On Free, Plus and Pro, OpenAI may use your content to train its models unless you turn off “Improve the model for everyone”; even then, a thumbs up or down lets it use the whole conversation. Business, Enterprise and Edu aren’t used for training by default, and a DPA is available. EU data residency is limited to eligible new Enterprise and Edu customers and to the API: ChatGPT Business doesn’t get it. More detail: ChatGPT and GDPR: can you put client data into it?

Claude (Anthropic)

On Free, Pro and Max, training hinges on a single setting, which also sets retention: if it allows your chats to be used, Anthropic may keep them for up to 5 years in de-identified form in its training pipelines. On Team and Enterprise, Anthropic acts as a processor, the DPA is built into the Commercial Terms and conversations aren’t used for training. For its commercial plans, Anthropic says data is stored in the US; Enterprise lets you set a retention period of at least 30 days. For Claude Cowork, the agent that reads whole folders: Claude Cowork for lawyers.

Mistral Le Chat (now Vibe)

Le Chat is now called Vibe, at the same address. Mistral hosts data in the European Union by default, with possible temporary transfers to subprocessors depending on the feature. But the French vendor trains on Free, Pro and Team by default unless you opt out (the admin does it for Team); only Enterprise is excluded by default. Mistral publishes no retention period for conversations: they’re removed after manual deletion, “typically after a short grace period”. More detail: Mistral Le Chat and privacy.

Microsoft 365 Copilot

With a work account, Microsoft doesn’t use prompts, responses or Microsoft Graph data to train foundation models, and interactions are stored with the firm’s other Microsoft 365 content, in a hidden folder of the user’s mailbox. How long they live depends on the Purview retention policies the admin sets. For European customers, Copilot is an EU Data Boundary service, with two exceptions: the Anthropic models offered in Copilot (off by default in the EU) and web search queries sent to Bing. Copilot’s specific risk lies elsewhere: it finds everything each colleague can open. More detail: Microsoft 365 Copilot and professional secrecy.

Which points are easy to miss?

A court-ordered hold can override your settings

In the copyright suit the New York Times brought against OpenAI, a preservation order required OpenAI, in 2025, to keep output log data that would otherwise have been deleted. According to OpenAI, it covered ChatGPT Free, Plus, Pro and Team and the API without zero data retention, but not ChatGPT Enterprise or Edu. An order of October 9, 2025 (case No. 23-cv-11195, U.S. District Court for the Southern District of New York) ended that obligation as of September 26, 2025. But OpenAI still keeps the logs already set aside, except those of users in the EEA, Switzerland and the UK, and it keeps preserving logs of accounts tied to domains named by the plaintiffs, wherever they come from.

The lesson goes beyond OpenAI: every policy cited here reserves retention “required by law”. A lawsuit you’re not a party to can freeze your conversations, and deleting them won’t help.

Deleting doesn’t mean erasing right away

OpenAI and Anthropic erase a deleted conversation within 30 days. Anthropic keeps a chat flagged by its safety systems for up to 2 years, and a chat someone rated with a thumbs up or down for 5 years, Team and Enterprise included. On a shared workstation, one colleague’s rating commits the whole exchange.

Connectors and web search fall outside the commitments

Commitments cover what the vendor processes itself. At Microsoft, web search queries sent to Bing aren’t covered by the DPA or the EU Data Boundary, and Microsoft acts as controller for them. At OpenAI, data residency doesn’t cover what goes through apps, MCP servers or web search. At Anthropic, consumer training data excludes raw connector content (Google Drive, for example), but not what’s copied into the conversation. Connecting the firm’s email or Drive widens what the AI can read, and what leaves.

Share links open the conversation to anyone who gets them

On a personal ChatGPT account, anyone with the link can see the shared conversation, files included depending on the share. On Claude Free, Pro or Max, a public link shows a snapshot of the chat, without attached files. At Mistral, anyone who receives the link, including by forwarding, can see the conversation. On ChatGPT Business, Enterprise and Edu, and on Claude Team and Enterprise, links only open for members of the organization; ChatGPT Business can’t disable links workspace-wide, unlike Enterprise.

An agent reads more than you paste

Agents change the scale. Claude Cowork reads connected folders and, since October 6, 2026 on Pro and Max, processes the local files it opens on Anthropic’s servers. Copilot draws on everything the user can open in Microsoft 365. It’s no longer an excerpt that leaves, it can be the whole file.

What does this mean for a professional bound by secrecy?

The table answers GDPR questions: who is the processor, for how long, where. Professional secrecy asks a different one. In France, Article 226-13 of the Criminal Code punishes the disclosure of secret information by the person entrusted with it with one year of imprisonment and a €15,000 fine. The text makes no distinction based on the contract with the recipient or on where its servers are. A DPA makes the vendor a processor under the GDPR; it doesn’t make it your client’s confidant.

For lawyers, the CNB (France’s national bar council) guide of April 2026 draws the rule: never share data about files or clients with generative AI, and pseudonymize or anonymize systematically before any use. It adds that a promise not to reuse data is “generally neither provable nor verifiable” (our translation) and, on sovereignty, looks for hosting by a company not subject to extraterritorial laws, which for now rules out American companies (see the CNB guide explained). The same reasoning applies to notaries, chartered accountants and the other professions bound by secrecy: generative AI and professional secrecy, profession by profession.

In practice, the table helps you pick the least risky tool for what isn’t secret, and rule out personal accounts. It doesn’t let you pick a “safe” tool for the file in clear text: no row gets there.

Which workflow keeps the file on your computer?

AI needs no names to summarize, compare or draft. The workflow that keeps the file on your machine:

1. Pseudonymize on your computer. Names, organizations, addresses, emails, phone numbers, IBANs and ID numbers are replaced with stable labels, such as [PERSONNE_1], across every document in the folder. With AnonymFiles, detection runs on your computer, with no account and no telemetry.
2. Review the copy next to the original. Look for indirect identifiers: an amount, a date, a reference, a project name. Nothing is exported without your approval.
3. Send the copy to the tool of your choice. A business plan with a DPA is still preferable, but what leaves no longer identifies anyone.
4. Put the names back on your computer. The app restores the real names in the AI’s answer, as text, Word or PDF. The encrypted mapping table never leaves the computer.
5. Check. Figures, dates, references and case law need checking, whatever the AI.

Further: pseudonymize a document for AI, then put the real names back, the mapping table, how AnonymFiles works and pricing.