AI and privacy
For each ChatGPT, Claude and Mistral Vibe plan: what happens to your input, how long it stays and whether an agreement is available. October 2026.
For work without personal data, any plan can be used if your own policy allows it. Once input contains client or employee data, it starts with ChatGPT Business: only there is there a data processing agreement and is your input not used for training by default. Only Enterprise offers European storage and your own retention periods.
This article explores one question; the full overview is on AI chatbots and the GDPR.
This article describes the position in October 2026, based on providers' public documentation. Plans, names and terms change several times a year: check the linked pages before recording a decision. We leave out prices; they appear on the providers' pricing pages and differ by country.
“May” here concerns the GDPR: what must be arranged before personal data goes to a provider. What that data includes is covered in May you put client data in ChatGPT, Claude or Mistral Vibe?. The wider explanation appears on ChatGPT, Claude, Mistral Vibe and the GDPR.
The difference is not the model. It is who decides what happens to the data.
On a personal plan, an individual contracts with the provider under consumer terms. Within those terms, the provider decides for itself what to do with the input. Your organisation is not a party: it cannot enforce anything, inspect anything or have anything deleted.
On a business plan, the organisation is the customer. It gets a workspace, administrators and commercial terms. The provider processes input on its behalf, which is what Article 28 GDPR requires a data processing agreement for. What to check in such an agreement is covered in Data processing agreement with OpenAI, Anthropic, Microsoft or Mistral: what should you check?.
Plan names change. Three changes worth knowing, as at October 2026:
For GDPR purposes, these four plans are the same. They differ in limits and models, not in their data arrangements.
In its explanation of how data is used to improve model performance, OpenAI says content from services for individual users may be used to train models. Users can turn this off with “Improve the model for everyone” under data controls. It then applies to new conversations.
Two things remain. A thumbs up or down may still make the whole conversation available for training. And the setting belongs to the employee's account: as an employer, you cannot see its state.
Ordinary conversations remain until the user deletes them. After deletion, OpenAI's page on chat and file retention says it schedules permanent deletion within thirty days, with exceptions for security and legal obligations. A temporary chat does not appear in history and is not used for training, but OpenAI may keep a copy for up to thirty days for safety.
A personal plan, including a paid one, is therefore no basis for processing client data. It is also where most unseen use takes place; see Shadow AI: what it is, how serious the risk is and how to control it.
ChatGPT Business is the plan for smaller organisations that you purchase yourself, without a sales conversation. OpenAI says a workspace requires at least two seats.
On its Enterprise privacy page, OpenAI says data from Business, Enterprise and the API is not used to train models by default. The same page says OpenAI can enter into a data processing agreement for these services. You request it using a form. It is therefore not in place automatically.
OpenAI says Business workspace administrators can view, open, export and delete users' conversations. Tell employees beforehand. At OpenAI, access to stored conversations is limited to authorised staff for support, investigating misuse and legal obligations, and to external parties assessing misuse under confidentiality obligations.
Enterprise is a contract you enter into with OpenAI's sales department. It adds three things for GDPR purposes.
Administrators decide how long data remains in the workspace. Deleted conversations leave the systems within thirty days, unless OpenAI is legally required to keep them. Administrators also receive a conversation log through a separate connection for oversight.
New Enterprise workspaces can choose where content is stored; Europe (the EEA and Switzerland) is one of the regions. This covers conversations, files and backups. For eligible customers, the model's computation can also remain in Europe.
OpenAI itself explains what this does not mean. Account information, billing data and metadata are excluded. Extracting text from a PDF, routing a request and signing in may take place outside the region. Anything sent to an external app through a connection falls under that party's terms.
As at October 2026, based on the OpenAI pages linked above. Check them before deciding.
| Plan | Training on input | Retention period | Data processing agreement | Administration and SSO | Storage location |
|---|---|---|---|---|---|
| Free | Possible until the user switches it off | Until the user deletes; then removed within 30 days, with exceptions | No | None | No choice |
| Go | Possible until the user switches it off; may include advertisements | As Free | No | None | No choice |
| Plus | Possible until the user switches it off | As Free | No | None | No choice |
| Pro | Possible until the user switches it off | As Free | No | None | No choice |
| Business (formerly Team) | Not by default | Deleted content removed within 30 days; check administrator settings | Yes, on request | Workspace and administrators, from two seats; check SSO | No choice |
| Enterprise | Not by default | Administrators set the period; deleted content removed within 30 days | Yes, on request | Workspace, SAML SSO, oversight log | Choice of regions, including Europe; not everything is covered |
The other two providers have a similar division, with their own differences.
For Claude Free, Pro and Max, the user chooses whether conversations may be used to improve the model, according to Anthropic's privacy centre under use for model training. That choice affects retention. Its page about consumer retention periods says a deleted conversation leaves systems within thirty days, but data from people allowing training may remain in the training environment for up to five years, disassociated from the user.
For Team, Enterprise and the API, Anthropic says conversations are not used for training, unless the customer expressly opts in. The data processing agreement need not be requested separately: Anthropic says it automatically forms part of the commercial terms, including standard contractual clauses.
Storage is the point to watch for anyone wanting to remain in Europe. On its page about server locations, Anthropic says data is stored in the United States. According to the Claude pricing page, Team has SSO sign-in, while audit logs and custom retention periods belong to Enterprise.
For Mistral Vibe, storage is the strong point: Mistral says data is hosted in the European Union by default. Depending on the feature, data may temporarily leave the EU through sub-processors named by Mistral.
Training follows the same pattern. The page about opting out of training says individual users are not opted out by default, and uploaded documents count as input. Enterprise customers are opted out by default. For Team, an administrator can switch off training for the entire organisation; the page does not state the default. Check the switch yourself.
For the data processing agreement, Mistral refers to its legal terms; read there which plans it covers. More about this provider appears in Mistral Vibe and the GDPR: is a European chatbot automatically safe?.
| Topic | ChatGPT | Claude | Mistral Vibe |
|---|---|---|---|
| Personal plans | Free, Go, Plus, Pro | Free, Pro, Max | Free, Pro, Education |
| Training on personal plans | Possible until the user switches it off | The user chooses | On until the user switches it off |
| First plan with a data processing agreement | Business | Team | Team; check the agreement |
| Training on business plans | Not by default | Not by default | Enterprise not by default; check Team yourself |
| European storage | Only Enterprise, as an option | Storage in the US | In the EU by default |
As at October 2026. If you use Microsoft 365, Copilot is a fourth option with its own question: what the assistant can see in your files. That is covered in Microsoft 365 Copilot and the GDPR: what can Copilot see in your SharePoint?.
A business plan arranges the relationship with the provider. It does not arrange four things.
PrivaSearch does not replace an AI assistant subscription. It addresses the remaining gap described above: what is entered. Instead of uploading a case file, you let the assistant search passages from which personal data has already been removed.
This happens in a fixed order. Documents are analysed inside PrivaSearch itself, without an external AI provider. Detected data is replaced with pseudonyms that remain consistent within a project. A colleague checks the result alongside the original and approves it; until then, a document is not searchable. That step is mandatory because automatic detection never finds everything. See How PrivaSearch works.
You can then search inside PrivaSearch or through a connected AI assistant. The connection works with Claude, ChatGPT Business and Mistral Vibe. The assistant receives only released, pseudonymised passages with a source reference, and no files. A business plan remains necessary: what the assistant receives becomes part of the conversation, and what you type yourself goes straight to the provider. Do not type names in the question.
Pseudonymised data remains personal data for your own organisation. Choosing the plan, keeping the record and carrying out the DPIA therefore remain your responsibility. Also see The risks of AI for your client files and the pricing page.
To see how this works alongside your current plan, request a demo.
For work without personal data, you may if your employer allows it. Not for client or employee data: Plus is a personal plan without a data processing agreement, and input may be used to train models until the user switches that off. The organisation has no administrative control over the account.
There is no difference: it is the same plan under a new name. OpenAI renamed ChatGPT Team to ChatGPT Business on 29 August 2025 and says renaming did not change the plan. If you still see Team in an article or quotation, it means Business.
That depends on the plan. On Free, Go, Plus and Pro, OpenAI may use your conversations to train models until you switch off “Improve the model for everyone”. On Business and Enterprise, this does not happen by default. On a personal plan, a thumbs up or down may still make the conversation available.
Only if your organisation has ChatGPT Enterprise or Edu and chose Europe when creating the workspace. Conversations and files are then stored in Europe. Account data and metadata are excluded, and some processing steps may happen elsewhere. According to the October 2026 documentation, personal plans and Business offer no choice of storage location.
All three offer a workspace and commercial terms. Claude Team does not train on input by default and includes the data processing agreement in its terms, but stores data in the United States. Mistral Vibe hosts in the EU by default; check its training switch yourself. The right choice depends on where your data may be stored.
No. A business plan arranges the agreements with the provider. The rest remains your responsibility: a lawful basis, an entry in your record of processing activities, justification for transfers, a DPIA where necessary, and agreements with staff about what they enter. No product arranges this for you. A plan is a prerequisite, not an endpoint.
Where an uploaded file ends up, how long it stays, who can access it and what deletion means. With four approaches for files containing personal data.
How schools use AI, why pupil data receives extra protection, which applications are prohibited or high risk, and what a school needs to arrange.
What ChatGPT stores, who can access it, how to turn off training, and what temporary chats, memory and shared links do. Includes a practical checklist.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →