Personal account or business account: the threshold that really matters for AI and data protection

Almost every conversation with a leadership team reaches the same question: Are we even allowed to do this? Usually it is about ChatGPT, Claude or Copilot with customer data, CVs or documents under NDA. The answers in circulation range from “only with consent” to “only with EU hosting” to “only self-hosted”.

Most of them focus on the wrong thing. For the large majority of companies one threshold decides: whether AI is used through a personal or consumer account, or through a business service with a data processing agreement (DPA) and no model training.

The account decides, not the model

With personal accounts the AI provider is a controller in its own right. That is stated in the privacy policies, for example of OpenAI Ireland for ChatGPT or Anthropic Ireland for Claude. Training on your inputs is the default there, with an opt-out. An opt-out does not change the provider’s role and does not replace a DPA.

In a business workspace the picture is different. There the provider is a processor: it processes on your instructions, the DPA is part of the contract, and your content is not used for training by default. This applies to OpenAI, Anthropic, Microsoft, Google and Mistral alike.

That gives a simple rule of thumb:

• Personal account (tier 0): only content without personal data and without confidentiality. Customer, patient and employee data do not belong there.
• Business workspace with DPA (tier 1): the standard recommendation. Personal data is fine to the extent the purpose requires it.
• API with EU processing, EU provider, self-hosting (tiers 2–4): normally questions of risk, trust and integration, not a precondition for lawfulness.

No consent needed for the processor

A common misconception is that customers or employees must consent to the use of AI. Using a processor requires neither an additional legal basis nor consent. The legal basis of the actual purpose carries it, for example performance of a contract when answering a customer enquiry. The German data protection conference and the Austrian data protection authority agree on this.

The reverse holds too: AI does not create a legal basis. What may not be processed without AI may not be processed with AI either. And a clause in the terms and conditions meant to allow “any AI use” is not explicit consent. “Could be relevant for the result” is not the same as “necessary”.

Where it really gets stricter

Three kinds of data deserve their own review:

1. Special categories under Art. 9 GDPR – health, religion, trade union membership, biometrics. An exception under Art. 9(2) is also needed for exactly this purpose. The term is broad: even a booking at a doctor’s practice can contain health data.
2. Confidential third-party data under NDA – a contractual question alongside the GDPR. A business service without training is functionally a cloud provider like email or file storage; most NDAs cover that. A consumer account with training does not.
3. Professional secrecy – patient files, legal case files and therapy notes are additionally subject to statutory confidentiality.

The real leaks are add-on features

In a business workspace the risk rarely sits in the chat itself but in the features around it: apps, web search, connectors and agents may pass data on under third-party terms or fall outside the contractual commitments. Configure these features deliberately before roll-out, not afterwards.

EU hosting is optional, not mandatory

Transfers to the USA are permitted if Chapter V of the GDPR is complied with. Microsoft, Google and AWS are certified under the EU-US Data Privacy Framework; in July 2026 the Austrian data protection authority confirmed that transfers to DPF-listed recipients are covered by the adequacy decision. OpenAI and Anthropic transfer on the basis of standard contractual clauses.

EU processing can still make sense, for example to shorten discussions with customers or the works council. It is then a deliberate decision about risk and trust, not a precondition for lawfulness.

What changed in August 2026

• Disclosure: since 2 August 2026 people must be told that they are interacting with an AI (Art. 50 AI Act). For a voice agent that means spoken, at the start. Where the AI is hosted belongs in the written privacy notice, not in the conversation.
• AI literacy: following the Digital Omnibus Regulation, Art. 4 AI Act requires measures to support staff AI literacy, but no guaranteed level and no certificates. The Austrian regulator RTR recommends documenting training.

What this means for you

If your people already use AI, the most important measure is not a ban but a good company account with a clear data rule: a business workspace with a DPA and no training, a simple traffic light for which data class may go into which tool, deliberately configured add-on features and documented training.

That answers most questions before they turn into a project. The rest – Art. 9 data, NDA customers, voice agents – you address specifically, with your legal counsel and clear criteria.

---

Sources (selection): OpenAI – How your data is used to improve model performance · Anthropic – Privacy Policy · Anthropic – Data Processing Addendum · DSK – Short paper no. 13 on processing · Austrian DPA decision of 31 July 2026 (RIS) · European Commission – Transparency obligations under Art. 50 AI Act · Regulation (EU) 2026/1744 (EUR-Lex) · RTR AI service desk – FAQ