AI and data protection: can your company knowledge go into Copilot or Claude?
Yes, if you use the business version, have a data processing agreement (DPA) in place and your permissions are clean. Free accounts with a private sign-in don't qualify. For highly sensitive knowledge, a model on your own hardware is the safer choice.
What data protection with AI comes down to
Three questions decide almost everything: Is the provider training on your inputs? Where is the data stored and processed? Is there a DPA under Art. 28 GDPR? According to the German data protection authorities (DSK), cloud AI often counts as processing on your behalf, and then such an agreement is mandatory.
What the providers commit to
As of November 2026.
No training
With Microsoft 365 Copilot, prompts, responses and company data are not used to train the models. Copilot only shows what the user is allowed to see anyway.
microsoft.com (opens in new tab)EU Data Boundary
For EU customers, Copilot runs inside the EU Data Boundary. The exception is Claude models in Copilot: off by default for EU tenants and not covered by the boundary.
microsoft.com (opens in new tab)No training
By default, Anthropic does not use inputs from Claude for Work and the API for training. The DPA with standard contractual clauses is automatically part of the commercial terms.
claude.com (opens in new tab)Stored in the US
By default, Claude data is stored in the US. Processing in an EU country is available through AWS Bedrock or Google Vertex.
claude.com (opens in new tab)
Personal account or business version?
If your company mostly sits on the right, company knowledge must not go into the AI. Switch first, then use it.
Fits: Microsoft 365 Copilot with a work account
Not for company knowledge: Free Copilot with a personal Microsoft account, conversations are used for training by default
Fits: Claude Team, Enterprise or API
Not for company knowledge: Claude Free, Pro or Max, each user decides about training
Fits: Work accounts managed by the company
Not for company knowledge: Employees use their own accounts
Fits: Training and history are off by default
Not for company knowledge: Training and history stay on because nobody checks
Fits: There is a list of approved tools
Not for company knowledge: Nobody knows which tools are in use
The bigger risk is permissions
Copilot finds everything an employee is allowed to see. That sounds safe, as long as the permissions are right. In many companies, though, every folder is shared with everyone, because that was convenient at some point. Nobody noticed before, because nobody searched through everything. An AI does exactly that. That is why Microsoft itself recommends reviewing and cleaning up permissions, for example in OneDrive and SharePoint, before you start.
It's similar with your own knowledge base for AI. According to the DSK, access can be controlled in the database through roles and rights, but not in the language model itself. So the rights concept belongs at the source.
What if the data must not leave the building?
Then the model runs on your own hardware. The DSK explicitly calls technically closed systems preferable. Open models with licences for commercial use are available now, and a well-equipped machine is enough to start. We will show what that looks like in practice on a page of its own soon.
Before company knowledge goes into an AI
Choose the business version
No accounts with a private sign-in, only work accounts you manage.
Check the DPA
With Microsoft and Anthropic it is part of the commercial terms. Read what it covers anyway.
Clarify where data is stored
EU Data Boundary, US storage or your own hardware, depending on how sensitive the knowledge is.
Clean up permissions
The AI sees what the user sees.
Bring in your data protection officer
According to the DSK, a data protection impact assessment is often required.
AI is rarely the data protection problem. Mostly it just shows what was already open.
What you usually ask us about AI and data protection.
The business versions bring the basics: no training on your data and a data processing agreement. The free version is a different story. Details on Business, Enterprise and the API are on our page about ChatGPT for companies. In the end, what matters is how you use the tool in your company.
Usually yes. If a provider processes personal data on your behalf, Art. 28 GDPR requires a data processing agreement. With Microsoft and Anthropic, it is part of the terms for business customers. Free accounts with a private sign-in don't come with one.
Not with Claude Team, Enterprise or the API. By default, Anthropic does not use these inputs for training, unless you explicitly send feedback. On the personal plans Free, Pro and Max, each user decides for themselves. For company knowledge, only the business plans are an option.
Microsoft 365 Copilot with a work account does not train on your data, runs inside the EU Data Boundary for EU customers and is covered by Microsoft's data processing agreement. The exception is the Claude models inside Copilot. Free Copilot with a personal account is not meant for company knowledge.
You want to use AI, but nobody dares to touch data protection? Let's talk.
We work out with you which tool fits your knowledge and clean up the permissions. Afterwards you know what can go into which AI. Not a traditional consultancy: we implement it with you. This doesn't replace legal advice, and we bring your data protection officer in.
Further reading

ChatGPT Business: Costs, Data Privacy, Limits
ChatGPT Business, Enterprise or API? What the plans cost, what happens to your data and when a licence is enough. Sorted out without vendor bias.
What Is RAG? Do You Still Need a RAG System?
What is RAG? Explained simply using a company knowledge base. Plus an honest answer on whether you still need a RAG system or MCP and APIs do the job.