GDPR and AI Integration: What Companies Must Get Right, and Where the EU AI Act Stands in November 2026
The GDPR does not treat AI as a special case. An assistant, an agent or a call to an LLM API is processing of personal data whenever personal data is in the prompt, in the documents the model reads or in the logs. What is new is where that data travels: into prompts, retrieval indexes, traces and a model provider's infrastructure. On top of that, the EU AI Act has applied in general since 2 August 2026. GDPR compliant AI integration therefore starts with knowing the data flows and designing them on purpose.
This guide covers the points that decide most projects: lawful basis, the agreement with the model provider, international transfers, the DPIA, and what the EU AI Act means for companies that use AI without building models. It is general information written from an engineering point of view. It is not legal advice; have your data protection officer or lawyer confirm the decisions for your case.
Where personal data shows up in an AI integration
- Prompts and attachments: whatever employees or customers type, paste or upload.
- Retrieved context: CRM records, tickets, e-mails or HR files that the system looks up and adds to the prompt.
- Outputs: a summary or a score about a person is personal data too.
- Logs and traces: stored prompts and answers, evaluation sets, debugging data.
- Fine-tuning data, if you train or adapt a model.
Map these first. An integration that touches no personal data, such as code generation or product descriptions, has little to do under the GDPR. One that reads customer tickets has all of the following to do.
GDPR and AI integration: lawful basis and purpose
Every purpose needs one of the legal bases in Article 6: contract, legal obligation, legitimate interests, consent and the others. Adding AI to an existing process does not inherit the old basis automatically. Check that the new use is compatible with the purpose the data was collected for. If you rely on legitimate interests, write down the balancing test. Special categories such as health data, biometric data or trade union membership also need a condition from Article 9; the simplest rule is to keep them out of prompts unless a condition clearly applies.
The European Data Protection Board addressed AI models in Opinion 28/2024 of December 2024. In short: legitimate interests can serve as a basis if the usual three-step test is passed, a model trained on personal data is not automatically anonymous, and a company that deploys a model should be able to show that it looked at how the model was built.
Controller, processor and the agreement with the model provider
In most integrations your company is the controller and the provider of the API, for example Anthropic, OpenAI or a cloud platform, processes the traffic on your behalf. That requires a data processing agreement under Article 28. Read the current terms for five points: whether inputs are used to train models, how long prompts are kept, which sub-processors are involved, where processing takes place, and how deletion and data subject requests are supported.
A consumer chatbot used with a private account is a different situation. Your company has no agreement there, so an employee who pastes customer data into it discloses that data without a contract. See also why shared AI chats are public.
International transfers
If the provider processes data outside the EEA, Chapter V applies. You need an adequacy decision for the destination or appropriate safeguards such as standard contractual clauses, together with a transfer assessment. For the United States, the EU–US Data Privacy Framework covers companies that are certified under it. The General Court dismissed a challenge to the framework in September 2025; an appeal was pending at the time of writing, so keep standard contractual clauses in place as a fallback. Several providers and cloud platforms offer processing in EU regions. That reduces the transfer question, but check support access and sub-processors before you treat it as closed. The Claude API integration guide lists the access routes.
DPIA, minimisation, retention and transparency
- DPIA (Article 35): required where a high risk is likely, for example systematic evaluation of people, monitoring of employees or special categories at scale. Run the screening early; it is cheaper before the architecture is fixed.
- Data minimisation: send the model only what the task needs. Mask or pseudonymise identifiers before the prompt. Pseudonymised data is still personal data, but the risk is lower.
- Storage limitation: set a retention period for prompt logs and enforce it technically.
- Record of processing (Article 30): add the AI use case and name the provider as a recipient.
- Information (Articles 13 and 14): update the privacy notice so people know AI is involved and who receives the data.
- Automated decisions (Article 22): no decision with legal or similarly significant effect may rest solely on automated processing unless one of the exceptions applies. Keep a person in the process who can change the result.
- Data subject rights: you must be able to find and delete a person's data in indexes and logs.
EU AI Act for companies: risk tiers and who owes what
The AI Act (Regulation (EU) 2024/1689) sorts AI systems into four tiers. Some practices are prohibited, for example emotion recognition in the workplace and in education outside medical or safety uses. High-risk systems include those used for recruitment and worker management, education, access to essential services such as credit, critical infrastructure and law enforcement. A third tier carries transparency duties. Everything else, which covers most internal assistants for drafting, search and coding, has no specific obligations beyond the general ones.
The Act separates the provider, who develops a system or puts it on the market under its own name, from the deployer, who uses it. A company that integrates a model through an API is normally a deployer. Two duties reach every deployer: the prohibitions, and AI literacy under Article 4, which since the 2026 amendment means taking measures to support the AI literacy of staff who work with the systems. For high-risk systems, Article 26 adds use according to the instructions, human oversight by competent people, monitoring, keeping logs for at least six months, informing workers and their representatives before workplace use, and informing people who are subject to decisions. The obligations for general-purpose models have applied since August 2025 and sit with the model providers.
If you build your own customer-facing chatbot and offer it under your name, you may count as the provider of that system. Either way, plan for the Article 50 rule that people must be told they are talking to an AI unless it is obvious.
AI Act timeline as of November 2026
- 1 August 2024: entry into force.
- 2 February 2025: prohibitions and AI literacy apply.
- 2 August 2025: governance rules and obligations for general-purpose AI models apply.
- 2 August 2026: general application, including the Article 50 transparency rules.
- 2 December 2026: generative systems placed on the market before 2 August 2026 must meet the marking duty for AI-generated content; the new prohibitions on non-consensual sexual deepfakes and child sexual abuse material apply.
- 2 December 2027: high-risk rules for Annex III systems (previously 2 August 2026).
- 2 August 2028: high-risk rules for AI in regulated products (previously 2 August 2027).
The last three dates come from Regulation (EU) 2026/1744, signed on 8 July 2026 and in force since 27 July 2026. Separate proposals to amend the GDPR itself were still in the legislative process at the time of writing; this article describes the rules in force. Fines under the AI Act reach EUR 35 million or 7 % of worldwide turnover for prohibited practices and EUR 15 million or 3 % for breaches of deployer and transparency duties. The GDPR's upper tier is EUR 20 million or 4 %.
How architecture reduces the legal burden
- Region: choose EU processing where the provider offers it.
- Retention: use zero-retention or short-retention options where they are offered and your use case qualifies.
- Pseudonymise before the prompt: replace names and numbers with tokens and map them back in your own system.
- Permissions in retrieval: the model should only see what the asking user may see.
- Data boundaries: define which classes of data may go to a hosted model and which stay inside, and enforce that in a gateway, not in a policy document.
- Local models: an open-weight model on your own servers has no external processor and no transfer. See the local LLM on-premise guide and private AI on your own servers.
Companies in Türkiye: the parallel KVKK regime
Companies in Türkiye face a similar set of questions under Law No. 6698 (KVKK), with its own rules on transfers abroad, standard contracts and registration, and guidance on generative AI from the Turkish authority. Companies that serve people in the EU can fall under both regimes. The Turkish article covers it: KVKK ve yapay zekâ.
Common mistakes
- Employees paste customer or staff data into public chatbots with private accounts.
- The company uses a model API without a data processing agreement.
- Prompts and answers with personal data are logged without an end date.
- The AI use case is missing from the record of processing and the privacy notice.
- The DPIA question is asked after the system is live.
- An HR or credit use case is built without checking the AI Act's high-risk list.
Checklist for an AI integration project
- Data flows mapped: prompts, retrieved context, outputs, logs
- Lawful basis and purpose documented for each use case
- Data processing agreement signed; training use, retention and sub-processors checked
- Transfer mechanism in place, or EU-only processing confirmed
- DPIA screening done and recorded
- Minimisation and pseudonymisation applied before the prompt
- Retention set for logs; deletion path tested
- Record of processing and privacy notice updated
- AI Act role and risk tier determined; staff trained
- Rules for employee use of AI tools written and communicated
Frequently asked questions
Is it GDPR compliant to use ChatGPT or Claude in a company?
It can be. The GDPR does not ban any model. You need a lawful basis for the personal data involved, a data processing agreement with the provider, a valid transfer mechanism if data leaves the EEA, and a retention rule for prompts and logs. A consumer chatbot used with a private account meets none of these.
Do we need a DPIA for an AI integration?
You need one when the processing is likely to result in a high risk to people, for example systematic evaluation of individuals, employee monitoring or special categories of data at scale. An assistant that only works on non-personal content usually does not. Document the screening decision either way.
Does the EU AI Act apply to a company that only uses an AI API?
Yes, as a deployer. The AI literacy duty and the prohibitions apply to every deployer. The heavier obligations apply only if the use case is high-risk, such as recruitment or credit scoring. The obligations for general-purpose models fall on the model provider.
When do the EU AI Act high-risk rules apply?
Regulation (EU) 2026/1744 moved them. High-risk systems listed in Annex III, such as those used in employment or for access to essential services, are covered from 2 December 2027. High-risk systems that are part of regulated products are covered from 2 August 2028.
Does a local model solve GDPR compliance?
No, but it removes two questions: there is no external processor for the model and no transfer to a provider. You still need a lawful basis, access control, a retention rule and a way to answer data subject requests.
Bottom line
GDPR and AI integration fit together when the data flows are known, the provider relationship is under contract and the architecture keeps personal data to the minimum. The AI Act adds a role and risk check that most internal assistants pass quickly and HR or credit use cases do not. This article is general information, not legal advice.
Planning an AI integration? Graf Clouds is an engineering company, not a law firm. We build integrations with hosted models such as Anthropic Claude and OpenAI with defined boundaries for what may leave the network, and we run open-weight models on customers' own servers when data has to stay inside. Your data protection officer and lawyers set the requirements; we implement them technically. See AI integration services, how to integrate AI into your company, or contact us.
Sources
- Regulation (EU) 2016/679 (GDPR), EUR-Lex
- Regulation (EU) 2024/1689 (AI Act), EUR-Lex
- Regulation (EU) 2026/1744 amending the AI Act, EUR-Lex
- European Commission: AI Act overview and application timeline
- EDPB: Opinion 28/2024 on data protection aspects of AI models
- European Commission: adequacy decisions
- KVKK: Personal Data Protection Law No. 6698 (English text)