# AI and the law: what GDPR and the AI Act ask of your team > Before staff paste work into a chatbot, two EU laws apply: GDPR for the data in the prompt, the AI Act for what the system is used for. What each asks of an IT team, with the dates as they stand in November 2026. Source: https://rfriedmann.de/blog/ai-and-the-law-gdpr-ai-act/ Published: 2026-11-08 · Track: learn · Level: Ops After "can we use ChatGPT for this?", the next question is usually whether it is allowed. In the EU two laws answer it, and they look at different things. The GDPR asks what is *in* the prompt. The AI Act asks what you *use the system for*. The two are often confused, and the inexpensive steps that are actually required get missed. This lesson is an overview: what each law asks of an ordinary IT department, which dates matter, and where to find the primary sources. It is not legal advice. Your data protection officer and your lawyer get the last word; this is so you walk into that meeting knowing the right questions. ## GDPR: follow the prompt The useful way to think about GDPR and LLMs is to follow one prompt with personal data in it and ask, at every stop, who holds it and on what legal basis.
One prompt with personal data, three stops, three sets of paperwork
[diagram omitted — see the page for the chart]
Each stop adds paperwork. A local model removes the processor and the transfer, not the GDPR itself: you still need a reason to process the data, and your own logs still hold it.
**Stop one: your side.** Pasting a customer email into a model is processing personal data. You need a legal basis under Art. 6 like for any other processing, a purpose, and as little data as the task needs. Often overlooked: prompt logs, chat histories and RAG indexes are personal data too, with retention periods and deletion requests (Art. 17) that reach them. **Stop two: the provider.** If the model runs at someone else's, that someone is normally your processor, and Art. 28 requires a data processing agreement (*Auftragsverarbeitungsvertrag*) before any personal data flows. Business and API plans usually come with one; a free consumer account an employee signed up for on their own does not. Three questions to put to every provider: is our data used for training, how long are prompts kept, and in which region are they processed. **Stop three: abroad.** Data leaving the EU needs a transfer mechanism under GDPR Chapter V. For US providers that is mostly the EU-US Data Privacy Framework, an adequacy decision from 10 July 2023, or standard contractual clauses. The General Court upheld the framework on 3 September 2025 (case T-553/23); an appeal is pending at the Court of Justice as C-703/25 P. The same court struck down the two predecessor frameworks, so treat the DPF as valid today and keep a plan B for the day it isn't. Two more checks belong here. A **data protection impact assessment** (Art. 35) is required when processing is likely to be high-risk for people, for example screening applicants, health data or monitoring staff. And **fine-tuning on personal data** is the hard case: you can delete a log line or re-index a document, but you cannot reliably delete a fact from model weights. That is one more reason [RAG beats fine-tuning for your documents](/blog/your-own-documents/). ## The AI Act: follow the use For the AI Act, which model you run matters less than your **role** and your **use case**. The use case then falls into one of four tiers: - **Prohibited** (Art. 5): social scoring, manipulative techniques that cause harm, emotion recognition at the workplace and in schools, and more. Banned outright. - **High-risk** (Art. 6, Annex III): areas where a wrong output hurts people, such as employment (CV screening, promotion decisions), education, credit scoring and critical infrastructure. Here deployers get concrete duties: human oversight, logs kept at least six months, and informing staff representatives before use at work (Art. 26). These duties apply from 2 December 2027. In Germany the works council already has a say today for any tool that can monitor staff (§ 87(1) no. 6 BetrVG). - **Transparency** (Art. 50): people must be told when they talk to a chatbot, deepfakes must be disclosed, and AI-generated text published to inform the public on matters of public interest must be labelled unless a human reviewed it and an editor takes responsibility. - **Minimal**: everything else, which includes most internal "help me draft this" use. No specific obligations beyond the general ones. The companies that make the models themselves, **general-purpose AI** providers, carry a separate set of duties (documentation, copyright policy, training data summary). Those duties fall on them, not on you, unless you train, or substantially modify, a model and release it. One duty does reach almost everyone: **AI literacy** (Art. 4). Providers and deployers must take measures so that staff using AI understand it well enough for their context. Regulation (EU) 2026/1744 softened the wording in July 2026: the duty is to *take measures*, and it explicitly does not require guaranteeing any particular level of literacy per person. In practice: a short, documented training before rollout. This course can serve as a starting point. ## The dates, as they stand in November 2026 The AI Act applies in stages. The "Digital Omnibus on AI", Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026, moved the high-risk dates back.
AI Act application dates after Regulation (EU) 2026/1744
[diagram omitted — see the page for the chart]
Before the Omnibus, Annex III high-risk rules were due on 2 August 2026 and the product rules on 2 August 2027. The new bans from 2 December 2026 cover AI that generates non-consensual sexual images of real people or child sexual abuse material. Providers whose generators were on the market before 2 August 2026 have until 2 December 2026 to mark AI output as such (Art. 50(2)).
The fines are substantial: up to 35 million euros or 7% of worldwide annual turnover for prohibited practices under the AI Act (Art. 99), up to 20 million euros or 4% for the most serious GDPR breaches (Art. 83). In Germany the Bundesnetzagentur is the market surveillance authority for the AI Act under the KI-MIG of July 2026; data protection stays with the federal and state data protection authorities. A second Omnibus, on the GDPR itself, is still being negotiated. Until it is adopted, the current GDPR text applies unchanged. ## Checklist

Before rollout

Stop and ask first

For most teams the summary is short. The left column is a week of work, not a project. The right column is where a [self-hosted model](/blog/self-hosting-an-llm-for-your-team/) or a [decision not to use AI at all](/blog/when-not-to-use-ai/) starts to make sense, and where your data protection officer should be in the room before anything is built. Who owns what the model produces, and whether it copies someone else's work, is a third law again: [copyright and AI output](/blog/who-owns-ai-output/). ## Primary sources - **GDPR**: Regulation (EU) 2016/679 on EUR-Lex. - **AI Act**: Regulation (EU) 2024/1689, plus the amending Regulation (EU) 2026/1744, both on EUR-Lex. The Commission's AI Act Service Desk explains it in plain language. - **AI models and personal data**: EDPB Opinion 28/2024, from the European Data Protection Board. - **Germany**: the Bundesnetzagentur for the AI Act, your state data protection authority for the GDPR. Laws change: the dates above already changed once this year. Check the consolidated text before you rely on a deadline.