It can be, on a business plan and with some data kept out. Free and personal plans, including paid ones like ChatGPT Plus or Claude Pro, may train models on your chats and let reviewers read them. Business plans don’t train on your data by default and come with the contract GDPR asks for.
So use a business plan with a data processing agreement, keep ID numbers, health details, card data and passwords out, and remove names the task doesn’t need.
Everything below is as of October 2026 and isn’t legal advice. Vendor terms change, so check the linked pages.
What happens to the text you paste in
A customer email you paste into an AI tool goes to the vendor’s servers. Four things decide the risk:
- Training: whether your text helps build the next model.
- Retention: how long the vendor keeps it, even after you delete the chat.
- Access: who at the vendor can read it.
- Location: which country it’s stored in.
On training, the European Data Protection Board treats a model trained on personal data as anonymous only if it’s very unlikely that anyone can extract that data through queries. OpenAI itself asks consumer users not to enter “sensitive information that you would not want reviewed or used”.
Paying for a plan doesn’t make it a business plan
ChatGPT Plus and Pro, Claude Pro and Max, and paid Google AI plans are all consumer plans. Business terms come with a business plan or a work account.
| Tool | Personal plans | Business plans |
|---|---|---|
| ChatGPT | Free, Plus, Pro: may train on chats unless you opt out | Business, Enterprise, API: no training by default |
| Claude | Free, Pro, Max: trains on chats if you allow it | Team, Enterprise, API: no training by default |
| Gemini | Personal account: trains while Keep Activity is on | Workspace account: no training or human review outside your organisation |
| Copilot | Personal Microsoft account: consumer terms | Work account: enterprise data protection |
- ChatGPT. You can opt out under Settings > Data controls, but a thumbs up or down can still put the whole conversation into training. Business plans aren’t used for training by default, and the API keeps inputs up to 30 days for abuse monitoring unless OpenAI approves zero data retention.
- Claude. Since 2025, personal users choose whether their chats train Claude. If they agree, Anthropic keeps de-identified data up to five years. Commercial plans aren’t used for training, and the API deletes inputs and outputs within 30 days.
- Gemini. With Keep Activity on, Google can train on your chats and have reviewers read some, keeping those up to three years. With it off, chats are still kept 72 hours. On a Workspace account, nobody reviews or trains on your content outside your organisation without permission.
- Copilot. With a work account, Microsoft processes your data for you under its data protection addendum and doesn’t train its foundation models on your prompts. A personal account runs on consumer terms.
Where the data is stored
GDPR restricts sending personal data outside the EU (Chapter V). On business plans:
- OpenAI stores data in Europe for eligible API customers and new ChatGPT Enterprise and Edu workspaces, but not for ChatGPT Business.
- Anthropic stores data in the US. Claude stays in the EU only through Amazon’s, Google’s or Microsoft’s cloud.
- Google Workspace admins on Business Standard and higher can keep data in Europe, Gemini chats included.
- Microsoft keeps EU customers’ business data inside its EU Data Boundary, except web searches and Anthropic models used in Copilot.
Data sent to the US relies on the EU-US Data Privacy Framework or on standard contract clauses. An EU court upheld the framework in 2025, an appeal is pending, and in July 2026 the EDPB asked the Commission to re-examine it after a US Supreme Court ruling on the independence of the FTC, one of its enforcers. It stays in force until repealed or annulled. Data kept in the EU depends far less on how that ends.
What GDPR asks of you
GDPR makes you the controller of your customers’ data, and the AI vendor processes it for you. Article 28 lets you use only a processor that gives sufficient guarantees, under a contract that says what it may do with the data, called a data processing agreement (DPA). All four vendors offer one on business plans: OpenAI, Anthropic, Google and Microsoft. Consumer plans come with a privacy policy instead.
Article 5 says data should be “adequate, relevant and limited to what is necessary”, so send the AI only what the task needs. Article 32 asks for security that fits the risk.
What should stay inside your own systems
Some data shouldn’t go into a chat tool at all, and into a workflow only with a clear reason and safeguards:
- Health and other special-category data. Health, religion, ethnic origin, sexual orientation, union membership, biometric and genetic data are prohibited by default under Article 9, with narrow exceptions. A 12-person dental clinic might use AI to draft appointment reminders, but treatment notes should stay in its patient system.
- ID numbers. Article 87 lets each country set its own rules for them. In Romania, Law 190/2018 requires safeguards, including a data protection officer, when a company processes ID numbers such as the CNP on the basis of its legitimate interest.
- Card and bank details. Anyone who stores, processes or transmits card data falls under the PCI DSS security standard, and a chat tool was never secured for card numbers.
- Passwords and access keys. On ChatGPT Business, workspace admins can read members’ chats, and on consumer plans vendor staff can in some cases. Anyone who can read the chat can read the password.
When a workflow we build needs this kind of data, we keep it in your own accounts and send the model only the fields a step needs, through a provider that doesn’t train on it. Sensitive work can run in the EU or on your own server.
Take out what the task doesn’t need
To sort enquiries, summarise complaints or draft a reply, the model needs the text of the message, not the sender’s phone number.
To pseudonymise data means to replace names and other identifiers with placeholders such as “Client 14”, and keep the list linking each placeholder to a real person in your own system. The vendor sees the content without knowing who it’s about. It’s still personal data for you, since you hold the list (Recital 26), but a leak at the vendor would reveal much less.
Checks to run this week
- List who uses which AI tool, on which account. Ask without blame. Look for client emails pasted into a personal ChatGPT.
- Check the plan. Free, Plus, Pro or Max means consumer terms. Move work onto a business plan or a work account.
- Check training. On personal accounts, look for “Improve the model for everyone” (ChatGPT), “Help improve our AI models” (Claude) or “Keep Activity” (Gemini).
- Check the feedback buttons. A thumbs up or down can share the whole conversation. Admins can turn rating off in ChatGPT workspaces and Claude Team and Enterprise.
- Check retention and region in the admin settings.
- Find the DPA, accept or sign it, and file a copy with your GDPR records.
- Write down what never goes in: health data, ID numbers, card details, passwords.
If a supplier builds AI into your processes, ask which AI provider gets your data, on which plan, kept for how long and stored where, and ask to see the DPA.
To see where personal data shows up in one of your processes, describe it in the free automation plan, which includes a note on personal data. Leave the names out.
Want to talk it through?
Book a call, or tell us about one process in a few lines.