Since ChatGPT was released, the use of large language models in German companies has grown rapidly. Their use brings significant data protection obligations that many responsible managers do not fully recognise. European data protection authorities clarified their positions further in 2026, with direct implications for everyday business.
Why ChatGPT can raise data protection issues
Using ChatGPT or similar services through cloud APIs from OpenAI, Anthropic or Google generally involves sending data to servers outside the EU. This may include not only direct user inputs but also metadata such as IP addresses, usage times and the frequency of particular requests. Under the GDPR, every transfer of personal data to a third country is a processing operation requiring its own legal basis.
The issue becomes more serious when employees knowingly or unknowingly enter personal data in prompts: customer names in an email to be summarised, applicant data being prepared for shortlisting or patient data included in a medical report. Each involves personal data processing for which the company is responsible as controller.
Planned and documented for GDPR compliance: data processing agreement, technical and organisational measures and records of processing activities.
The legal position: three central requirements
The German Data Protection Conference's guidance sets out three central requirements for companies using ChatGPT or similar services.
First, a legal basis for processing. Companies need a valid basis under Article 6(1) GDPR. For cloud APIs, Article 6(1)(f), legitimate interests, is generally considered, but only where the interests of data subjects do not override it. For sensitive data, such as health data under Article 9, this basis is generally unavailable.
Second, a data processing agreement. Companies using an AI service through an API must conclude an agreement with the provider under Article 28 GDPR. OpenAI has offered standard agreements since 2024, but these should be reviewed carefully, particularly the provisions on further processing of input data and subprocessors.
Third, a data protection impact assessment. Systematic LLM use affecting many employees or customers generally requires a DPIA under Article 35 GDPR. This applies particularly where AI supports automated decision-making, such as applicant shortlisting or creditworthiness assessment.
Technical alternatives: Private AI for compliance
Faced with regulatory complexity, more companies are choosing Private AI, operating models on their own servers or with a German hosting provider. Data remains within the European legal jurisdiction and the company retains full control of processing.
This approach is considerably more practical in 2026 than it was two years earlier. Open-source models such as Llama, Mistral and their successors achieve quality comparable to proprietary models in many use cases. More efficient hardware and optimised inference frameworks have reduced operating costs. A mid-sized company can now operate a capable language model on German servers without a six-figure budget.
Private AI is not a universal solution. GDPR requirements also apply to self-hosted models. Model training and use must be documented, technical and organisational measures must be appropriate and data subject rights must be respected. A language model trained on customer data, for example, must be able to implement deletion requests under Article 17 GDPR, a technically demanding requirement.
EU AI Act transparency obligations since 2 August 2026
The transparency obligations under Article 50 of the EU AI Act have applied since 2 August 2026, among other provisions. Depending on the system and role, they concern disclosure of interaction with AI and labelling certain AI-generated or manipulated content. For Article 50(2) obligations applying to existing systems, there is a transition period until 2 December 2026. Legal status: 6 November 2026. Source: Regulation (EU) 2026/1744
We clarify which EU AI Act obligations apply to your use case in our AI compliance consulting.
Companies should check which transparency obligations apply to their systems and how to implement them in operation. The Article 50(2) transition period for existing systems does not postpone all transparency obligations.
Practical steps for companies
The current legal position suggests five practical steps. First, create an internal AI tool policy clearly defining which data may be entered into which systems. Second, maintain a register of all AI applications, analogous to records of processing activities under Article 30 GDPR.
Third, involve the data protection officer early, ideally while evaluating new AI tools. Fourth, train employees regularly in both using the tools and handling personal data appropriately in AI contexts.
Fifth, companies seeking greater control can use Private AI infrastructure with processing on German servers. This does not remove all compliance requirements, but significantly reduces exposure and simplifies documentation for supervisory authorities.
Traceable documentation can be planned
Using ChatGPT and other LLMs in a company is not inherently unlawful. It requires careful planning, documentation and ongoing review. Companies making the right arrangements can use AI technologies while addressing regulatory risk. Acting without a strategy risks fines and the trust of customers and business partners.
The requirements are complex but manageable. With appropriate advice and a structured approach, GDPR-compliant AI use can be established within a few weeks.