Contents
- When the GDPR Applies to AI Features
- Six Checks Before Launch
- Architecture Options Compared from a Privacy Perspective
- Technical Measures: Privacy by Design for AI Features
- Data Subject Rights and Automated Decisions
- Think GDPR and the EU AI Act Together
- Checklist: A GDPR-Compliant AI Feature
- Conclusion
- Frequently Asked Questions
Building GDPR-compliant AI features is entirely achievable when data protection is part of the architecture from day one: a clear legal basis, a data processing agreement with the AI provider, as little personal data in the prompt as possible and transparent information for users. This guide covers the key checks and technical measures for chatbots, text analysis, image recognition and similar features in your app. It is not legal advice.
When the GDPR Applies to AI Features
The GDPR applies as soon as personal data is processed – any information relating to an identified or identifiable person. With AI features in apps, that happens more often than many teams expect:
- Users type names, addresses, health details or contract numbers into a chat.
- Uploaded documents or photos contain people, signatures or licence plates.
- Voice recordings are personal data in themselves, because a voice can be linked to a person.
- Personalisation and recommendations are based on the usage behaviour of an account.
Even if your feature "only summarises text", assume that personal data will end up in the input sooner or later – and design accordingly.
Six Checks Before Launch
- Define purpose and legal basis: what exactly is processed, and why? Common bases are performance of a contract (the AI feature is part of the service), legitimate interests or consent. Special categories such as health data come with stricter requirements.
- Clarify providers and roles: an external AI service is usually a processor. You need a data processing agreement (DPA) under Art. 28 GDPR. Our article on data processing agreements for app development and hosting covers the details.
- Check international transfers: if the provider sits outside the EU or has access from third countries, you need a transfer mechanism – such as an adequacy decision or standard contractual clauses. Many providers now offer processing in EU regions.
- Exclude training on your data: check contractually whether inputs are used to train the model and choose business terms that exclude this.
- Assess the need for a DPIA: large-scale processing of sensitive data, systematic evaluation of people or new high-risk technology may require a data protection impact assessment under Art. 35 GDPR.
- Inform and document: update the privacy policy and the record of processing activities, and set deletion periods.
Architecture Options Compared from a Privacy Perspective
Where the model runs determines much of the compliance effort. The common options:
| Option | Data protection | Effort and cost | Typical use |
|---|---|---|---|
| API of a major AI provider, EU region | Manageable with DPA, EU processing and no-training terms | Low entry cost, usage-based fees | Chat, summaries, classification |
| Model via a European cloud provider | Data stays in the EU, fewer transfer questions | Medium, model choice sometimes limited | Corporate data, public sector |
| Open-weight model on your own server | Maximum control, no disclosure to third parties | Higher: operations, GPU costs, updates | Highly sensitive data, high volumes |
| On-device model in the app | Data never leaves the device | Limited model size, device fragmentation | Text recognition, simple classification, offline features |
A combination often works best: sensitive pre-processing on the device or your own server, demanding language tasks via an API with pseudonymised data. Read more about server location in data storage in the EU and international transfers.
Technical Measures: Privacy by Design for AI Features
Art. 25 GDPR requires data protection by design. These measures have proven effective for AI features:
- Data minimisation in the prompt: send only the fields the task needs. Instead of the full customer record, the product category is often enough.
- Pseudonymisation: replace names, email addresses or IBANs with placeholders before sending, and resolve them again in the app.
- Input filters: advise users not to enter sensitive data and automatically detect obvious patterns such as card numbers.
- Short retention: keep chat histories and logs only as long as function and debugging require. Where providers offer zero-retention options, consider them.
- Access control in the knowledge base: in RAG setups the assistant may only draw on documents that the specific user is allowed to see.
- Encryption: TLS in transit, encrypted storage, documented key management.
Data Subject Rights and Automated Decisions
Users have rights of access, rectification and erasure – including for data created by AI features. Plan where chat histories, generated content and embeddings are stored so you can remove them completely when a deletion request arrives.
Decisions made solely by automated means that significantly affect people – such as a credit refusal or pre-screening job applicants – are particularly sensitive. Art. 22 GDPR sets tight limits here. In practice: the AI prepares, a human decides, and the people affected are told that AI was involved.
Think GDPR and the EU AI Act Together
Since 2024, the EU AI Act applies as well. It does not replace the GDPR but adds requirements for AI systems themselves: transparency when people interact with an AI, for instance, and stricter duties for high-risk applications. Obligations are phased in, many of them from August 2026. Clean GDPR documentation gives you a strong head start. See our overview of the EU AI Act for app providers.
Checklist: A GDPR-Compliant AI Feature
- Purpose, data categories and legal basis documented
- DPA signed with the AI provider and hosting partner
- Processing region and international transfers checked
- Training on customer data contractually excluded
- Data minimisation and pseudonymisation implemented
- Retention and deletion periods set for histories, logs and embeddings
- Need for a DPIA assessed and the outcome documented
- Privacy policy and in-app AI notice updated
- Human review planned for consequential decisions
A complete list for your whole app is available in our GDPR checklist for apps. We develop and host GDPR-compliantly in Germany – use the offer wizard to have your AI project analysed free of charge.
Agentic coding workshop for your team? From €799 net – request a quote by e-mailConclusion
Making AI features GDPR-compliant is mainly about deciding consciously which data flows where. With the right hosting option, a DPA, no-training terms, data minimisation and a clear deletion concept, most use cases can be delivered on solid legal ground. Teams that only check privacy after development usually pay twice – through rework and delayed launches.
Frequently Asked Questions
Can I use ChatGPT-style services in my app in a GDPR-compliant way?
Generally yes, provided there is a DPA, international transfers are covered, inputs are not used for training and users are informed transparently. Details depend on the provider and your use case.
Do AI features require a data protection impact assessment?
Not always. A DPIA is mainly needed for sensitive data, large-scale processing or systematic evaluation of people; assess and document the decision in every case.
Is a model on my own server automatically GDPR-compliant?
No. It resolves questions about disclosure to third parties, but legal basis, purpose limitation, deletion periods, access control and data subject rights still have to be implemented.
Do users need to know they are talking to an AI?
Yes. Transparency is advisable under the GDPR and the EU AI Act explicitly requires it for many systems that interact directly with people.