Using ChatGPT at work has become routine, often faster than data protection and compliance teams can react. The productivity gains are real: drafts, summaries, code, translations. But as soon as personal data, customer information, draft contracts or source code are pasted into an input field, a productivity question turns into a data protection question with concrete duties under the GDPR.
At a glance
- Free accounts and ChatGPT Plus may use inputs for model training by default. ChatGPT Business and Enterprise offer better contract terms and still do not settle the question of which jurisdiction applies.
- The US CLOUD Act obliges US providers to hand over data on an order from authorities, whether the servers are in the US or in Europe. No contract removes that.
- The biggest practical risk is shadow AI, meaning unmanaged use by employees.
What happens to your data
The plan makes the biggest difference. With free accounts and ChatGPT Plus, inputs may be used to improve the models by default, and the setting is on out of the box. An opt-out exists. The company, however, carries the responsibility for setting it correctly for every person, and one forgotten checkbox can amount to a data protection incident.
With ChatGPT Business and Enterprise, things look clearly better. Inputs are contractually not used for training, there are options for retention, admin functions and a data processing agreement. Two points remain. Data is still processed technically (inference, abuse detection, time-limited retention). And no plan changes the jurisdiction the provider is subject to.
CLOUD Act, Schrems II and the third-country transfer
Any processing of personal data by a US provider is a third-country transfer within the meaning of Art. 44 et seq. GDPR. Since the Schrems II ruling of the Court of Justice, pointing to standard contractual clauses is no longer enough. A review is needed, and where applicable additional safeguards. The EU-US Data Privacy Framework of 2023 eases the situation formally. It is legally contested and therefore offers real but provisional reliability.
On top of that sits the US CLOUD Act. It obliges US providers to hand over data in their possession, custody or control on an order from authorities, whether the servers are in the US or in Europe. What counts is where the provider is based. That reach cannot be fully excluded by contract. How location and control relate is explained in the article Data residency and sovereignty.
Contract, works council and shadow AI
Even if you accept that jurisdiction, duties remain. A data processing agreement under Art. 28 GDPR is a prerequisite, and plans for private individuals do not come with one. In Germany, co-determination often comes on top. If an AI tool can monitor employees or make their performance measurable, section 87 of the Works Constitution Act frequently applies in practice, and the works council has to be involved. A works agreement on AI use creates clarity for both sides.
Shadow AI arises when employees, with no official solution, use private accounts and enter confidential content. A survey among law firms shows how widespread this is. The 2026 benchmark report by Wolters Kluwer rests on the answers of 99 German law firms (out of 633 respondents across Europe). According to it, 63.3 percent of the German firms use AI in daily work. Of these, 82.5 percent use generic tools such as ChatGPT, without central control. The sample was recruited through the vendor itself and is therefore not a representative cross-section of all law firms. It still shows that use is real and widespread before a company has decided whether it should be allowed.
A blanket ban drives use underground. What helps is an official, secure alternative that is at least as good as the tool people use in secret.
The sovereign alternative
This is where a setup run by you or a partner comes in. Capable open language models run on your own servers or on servers hosted in Germany. Technically, inference software such as vLLM runs them, with a gateway in front that has an interface like the one cloud services use. Applications address it like a cloud service. With purely self-operated models and no external services, inputs do not leave your own environment. How this works technically is described in the article Running your own language models.
The productivity gain in a company rarely comes from the model alone. What decides it is access to your own knowledge through an evidence-based architecture that answers from your documents and cites the source. The article RAG with sources and the archive product RETRAYA show what that looks like. A self-hosted solution does not come for free. It needs infrastructure, operations and model maintenance. In return, with purely self-operated models the third-country transfer falls away, because the data does not leave your own environment.
What helps and what to watch for
What helps
- Clarify which data classes may go into an AI tool at all, and put it in a policy.
- Sign the data processing agreement with every provider and document the third-country question.
- Involve the works council and the data protection officer early.
- Choose a self-hosted environment for confidential, regulated or business-critical data.
What to watch for
- Personal and confidential data belong only in accounts covered by a data processing agreement.
- The EU-US Data Privacy Framework should not be the only and lasting legal basis, so plan alternatives alongside it.
- Shadow AI belongs in your rules on permitted tools.
- Servers in Europe do not automatically rule out access under the CLOUD Act.
The full reasoning behind the bigger picture is on the page about sovereign infrastructure.