Rules
When an AI provider is your processor, what the agreement must contain and where each provider publishes it, with a checklist of questions to ask.
An AI provider is your processor when it processes personal data for your organisation, on your instructions and for your purpose. This applies to business plans and the APIs of OpenAI, Anthropic, Microsoft and Mistral. A free or personal account uses consumer terms, and the provider decides for itself what it does with the input.
The distinction comes from GDPR definitions. The controller determines the purpose and means of processing; the processor acts on its behalf. This is in Article 4, and the requirements for the relationship between the two are in Article 28 GDPR. Having a client file summarised by an AI service your organisation purchases for business means engaging a processor.
The providers draw this boundary themselves. The first paragraph of OpenAI's services agreement says it applies to the API, ChatGPT Enterprise, ChatGPT Business and other services for businesses and developers, not services used by consumers or individuals. Anthropic's commercial terms and Mistral's similarly refer individuals to separate consumer terms.
An employee pasting a client file into a personal account therefore uses a service for which your organisation has no agreement. Which plans fall under business terms is set out per provider in which plan you may use for business.
Then it is not a processor for that part. Article 28(10) says that a processor determining purposes and means itself is considered the controller for that processing. Mistral writes in its agreement that it acts as controller for model training, detecting abuse and usage statistics. You therefore have no right to instruct it for that use, and must assess whether you may disclose your data for those purposes.
Article 28(3) first requires a description: the subject matter and duration of processing, its nature and purpose, the type of personal data, categories of data subjects and the controller's rights and obligations. Eight commitments for the processor follow.
The agreement must be in writing, including electronically (paragraph 9). Clicking “accept” on business terms incorporating the agreement may therefore suffice.
The table below gives the position as of October 2026, based on the providers' own documents. Its periods concern the contract; how long a conversation is retained during use is a product setting. Documents change: read the version applying on the day you sign or renew.
| Provider | Which plans | Where to find it | Point to check |
|---|---|---|---|
| OpenAI | The API, ChatGPT Business, ChatGPT Enterprise and other business services under the services agreement. Not ChatGPT Free, Plus or Pro. | The addendum on OpenAI's website, part of the services agreement. A form to have it signed is at the bottom. | No breach notification deadline in hours. OpenAI mentions configurable retention for Enterprise and Edu, not Business. |
| Anthropic | The API and Claude for organisations (Claude for Work) under commercial terms. Not Claude Free, Pro or Max. | The addendum on Anthropic's website, automatically part of the commercial terms. | When using Claude through another party's platform, that party's terms apply. |
| Microsoft | Copilot and Copilot Chat as used by organisations, under Microsoft 365 terms. | The products and services addendum on Microsoft's licensing website, in many languages. | Web searches fall outside the addendum. Anthropic models fall outside the European data boundary. |
| Mistral | Mistral Vibe (formerly Le Chat) and Studio for organisations, under commercial terms. The documents name no specific plan. | The addendum on Mistral's legal website, with the business customer documents. | Training is enabled by default for some products. Mistral is itself controller for training. |
OpenAI's addendum was updated on 1 December 2025 and applies from 1 January 2026. Customers in the European Economic Area enter into it with OpenAI Ireland Ltd. OpenAI describes itself as processor and processes customer data only to provide the service. The services agreement adds that OpenAI does not use customer content to develop or improve services unless the customer explicitly agrees.
On its enterprise privacy page, OpenAI says that with ChatGPT Enterprise, ChatGPT for Healthcare and ChatGPT Edu, you determine how long data is retained. ChatGPT Business is not named in that sentence. After the agreement ends, OpenAI deletes customer content within thirty days unless retention is required by law.
Anthropic's addendum forms part of its commercial terms. In its privacy centre, Anthropic explains that you do not sign it separately: accepting commercial terms means accepting the addendum. Customers in the European Economic Area contract with Anthropic Ireland, Limited. The commercial terms specify that Anthropic does not train models on customer content.
The addendum gives two firm deadlines: security breach notification within 48 hours at the latest, and deletion within thirty days after the agreement ends.
Microsoft uses one addendum for all its products and services; the latest version is dated 22 May 2026. Its documentation on enterprise data protection says organisations' use of Copilot and Copilot Chat falls under that addendum and the product terms, with Microsoft as processor, and that prompts and answers are not used to train foundation models.
The same documentation gives two exceptions. Searches Copilot sends to the web search service fall outside the addendum: Microsoft itself is controller there. And Anthropic models within Copilot run with Anthropic as Microsoft's sub-processor, but outside the European data boundary. They are therefore off by default for EU customers. For more about what Copilot sees of your files, see Copilot and SharePoint.
Mistral's addendum has applied since 27 July 2026 and forms part of the commercial terms. The contracting party is a French company, so transfers outside the EU do not arise in principle.
Training stands out. The commercial terms say Mistral does not use customer data to train models, except for products where training is on by default and the customer has not disabled it. Mistral's help centre names them: with Mistral Vibe, input and output are used for training by default; with Mistral Vibe for Enterprise, they are not by default. Mistral's assistant was previously called Le Chat; its own pages now call it Mistral Vibe. Why a European provider does not resolve the rest of the issue is explained in Mistral Vibe and the GDPR.
The table below compares the arrangements you rely on in an incident or termination. As of October 2026.
| Subject | OpenAI | Anthropic | Microsoft | Mistral |
|---|---|---|---|---|
| Objecting to a new sub-processor | Within 30 days of notice | Within 15 days | 6 months' prior notice; 30 days for AI sub-processors, with an option to disable them | Within 10 days |
| Breach notification | Without undue delay | Without undue delay, within 48 hours at the latest | Without undue delay, within 72 hours at the latest | Without undue delay |
| After the agreement ends | Deletion within 30 days | Return or deletion within 30 days | 90 days to retrieve data, then deletion within 90 days | No longer accessible after 30 days; deletion under its own policy |
| Audit | Information at most once a year; audit at the customer's expense | External auditors' reports on request; own audit subject to conditions | Audit reports in a portal; own audit by an external party at the customer's expense | Document-based audit; on-site at most annually, with 90 days' prior notice |
All four use prior general authorisation for a list they may change; the GDPR then requires prior notice and an opportunity to object. Ten or fifteen days is short if notice reaches an inbox nobody reads. Subscribe to notifications and record who assesses them. At OpenAI and Microsoft, the final outcome of an objection is that you may terminate the affected part.
OpenAI and Anthropic are US companies with Irish contracting entities. Their agreements base transfers on the standard contractual clauses the European Commission adopted on 4 June 2021; OpenAI also names an adequacy decision as a possible basis. Microsoft says it participates in the EU-US Data Privacy Framework and also applies standard contractual clauses.
That framework applies only to participating US companies. On 3 September 2025, the EU General Court dismissed a challenge to the decision, but an appeal is pending before the Court of Justice. Check whether the Court has since ruled, and record the basis for your transfer in your record of processing activities.
Keep the answers with the entry in your record of processing activities. How to incorporate this into a policy is explained in Creating an AI policy.
Three things remain.
The main explanation of chatbots and the GDPR, with differences between plans, is on ChatGPT, Claude, Mistral Vibe and the GDPR.
PrivaSearch sits between your documents and the AI model. Personal data is detected inside PrivaSearch and replaced with pseudonyms, and a colleague approves the result before a document becomes searchable. A connected assistant receives only released, pseudonymised passages with a source reference, never the original file. The steps are on how PrivaSearch works.
This does not remove the need for an agreement with your assistant's provider. What PrivaSearch returns becomes part of the conversation and reaches that provider. What you type into the assistant goes there directly, outside PrivaSearch. Pseudonymised data remains personal data for your organisation, because you hold the key. We therefore recommend connecting only a business account with a data processing agreement and recording the transfer in your record of processing activities. The connection works with Claude, ChatGPT Business and Mistral Vibe.
The same questions in the checklist above apply to PrivaSearch itself. Answers are in our data processing agreement and sub-processor overview. Hosting, storage and detection take place in data centres in the EU; the explanation is on hosting in the EU.
Your environment also creates an entry for the record of processing activities and a DPIA starter document: starter documents, not legal advice. To see how it works, book a demo.
Yes, as soon as employees process personal data in it for work. That agreement exists only for OpenAI's business services: the API, ChatGPT Business and ChatGPT Enterprise. ChatGPT Free, Plus and Pro use consumer terms and have no data processing agreement. On such an account, you therefore cannot meet the GDPR requirement that processing be governed by an agreement.
Usually not. At Anthropic, Microsoft and Mistral, the agreement forms part of the business terms you accept when subscribing. OpenAI also offers a form to have the addendum signed. The GDPR requires a written agreement and includes electronic form. Keep a copy of the version applying when you accepted, together with the date.
A processor may process data only on your instructions. OpenAI and Anthropic exclude training on business customer content, unless you explicitly agree to it at OpenAI. Microsoft does not use prompts and answers to train foundation models. At Mistral, training is on by default for some business products until you disable it. Check the setting in your administration environment, as of October 2026.
A sub-processor is a party the provider engages to process your data, for example for storage or computing power. This requires your authorisation. The four major providers request prior general authorisation and announce changes to a list. You retain the right to object, but the deadline varies from ten days to six months.
No. The agreement governs only the relationship with the provider. You also need a lawful basis, an entry in the record of processing activities, a transfer basis if data leaves the EU, and a DPIA for high risk. You must still process as little data as possible. No agreement takes care of an organisation's GDPR obligations for it.
For Copilot and Copilot Chat as used by organisations, yes: Microsoft describes itself as processor under its products and services addendum. Searches Copilot sends to the web fall outside it; Microsoft itself is controller for those. Anthropic models within Copilot are covered by the addendum but fall outside the European data boundary, and are off by default for EU customers.
The AI Act timeline after the digital omnibus: what applies already, what has moved to 2027 and 2028, and what AI chatbot users need to arrange now.
What a deepfake is, which rules apply now, which prohibition is coming and how organisations tackle fraud using a fabricated voice or video.
What the Dutch Algorithm Register and IAMA are, what is mandatory and what is not, and what a municipality arranges when staff use AI chatbots.
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 →