AI agents and GDPR: what the law asks of companies

AI agents break a goal into steps, pick tools and write into systems without a person approving each step. The Bavarian data protection supervisor has announced a focus on them for 2027, so companies should know what the GDPR asks of them.

Klaus Bauer4 min read

Illustration: A wind-up toy cart with a toolbox rolls toward a barrier that a lever beside the track can open.

What the supervisory authority has announced

The Bavarian State Office for Data Protection Supervision (BayLDA) plans to make AI agents a focus of its work in 2027 and to run a thematic series of audits on them. The Dr. Datenschutz article on AI agents reports this. The office describes agents as systems that increasingly process information on their own, merge several data sources and carry out actions.

The questions it names are familiar ones: purpose limitation, transparency, access permissions and effective human oversight. It also covers Art. 22 GDPR and the question of who is responsible. By its own account, the office sees preventive data protection as especially valuable while new technology is becoming established in companies.

Why agents change the way data protection is assessed

A classic AI system answers a request and stops. An agent decomposes a task, calls interfaces, reads documents and repeats this until the goal seems reached. The person sets the task and no longer controls every processing step.

With conventional software, each data flow can be described in advance. With an agent, the concrete path emerges at runtime, so the data actually processed depends on the route it chooses. The assessment therefore has to describe the space in which the agent may act. That means the sources and systems it can reach, the permissions it holds and the points where its limits are anchored technically.

Which GDPR duties apply to AI agents?

No separate data protection law exists for agents. The GDPR applies, and the AI Act sits alongside it as product and market regulation without replacing it. What changes is the effort needed to establish the facts. Four points structure the review:

  • Roles: who is controller and who is processor depends on who decides purpose and means. Model provider, platform operator and your company can hold different roles at once.
  • Purpose: a vague purpose such as process support does not suffice if the agent in practice pre-sorts applications or rates customer requests.
  • Legal basis: contract, legitimate interest, consent or legal obligation. The bar is higher for employee data and special categories of data.
  • Transparency and data subject rights: information duties, access, erasure and rectification must work across every component involved.

Art. 22 GDPR deserves particular attention. An agent that only proposes poses little problem while a person genuinely reviews the proposal. If proposals are routinely adopted unchanged, for example rejection drafts for applicants at high volume, the decision effectively lies with the system even though a human formally confirms.

What companies should check and set up now

The sources locate the biggest risks in how an agent is connected rather than in the model. Agents often inherit the rights of a technical user and then see more than the person who gave the task. Stored intermediate results, history data and vector indexes count as processing in their own right and need deletion periods. Calls to third-party services can amount to a transfer, even if the log shows only a technical call.

Documents, emails or web pages can also contain instructions that an agent mistakes for its task, which may lead to unintended data leaks. Logs are needed for traceability but often hold the very personal data that should be minimised. Assess these points together, because a broadly authorised agent with persistent memory and external access can quickly justify a data protection impact assessment. A practical start looks like this:

  1. List every agent with its purpose, permitted tools and data sources.
  2. Give each agent the narrowest set of rights and a responsible department, and treat it like a role.
  3. Define storage locations and deletion periods for intermediate results and logs.
  4. Record what the agent must not do.
  5. Check that the record of processing activities, processor contracts and the technical configuration show the same picture. If they differ, the configuration is usually what counts.

Background on the legal framework is on our page about GDPR-compliant AI, and the topic of rights ties in with security and data protection.

Building human approval so that it works

Oversight has to be feasible in daily work and cannot live only in the process description. If a reviewer approves hundreds of drafts an hour, the check exists on paper only. Limit approval to steps that write to systems or contact people, show the reviewer the sources behind a proposal, and keep volumes at a level that allows real reading. Measure how often proposals are changed or rejected, since a rate near zero is a warning sign.

Our agents with approval prepare tasks, and anything that writes to a system waits for a person to release it. The platform connects sources such as SharePoint or file servers read-only with existing permissions, and each answer cites its source. It runs hosted in Germany or on-premise, as described under local AI. Whichever tool you use, the documented limits and the approval step remain your own responsibility.

Sources

More articles

All articles

Run Text Generation Inference with Docker on Your Server

Text Generation Inference (TGI) serves open language models from a single Docker container on your own hardware. This tutorial starts it, sends requests over HTTP and keeps prompts inside your network.

Tutorial4 min

Cloud outage liability: who answers when Microsoft fails

On 12 November 2026 a software fault took Exchange Online and Outlook offline for about 36 hours. A legal analysis published shortly afterwards argues that responsibility for such outages often lies with the operator rather than the provider.

Regulation4 min

GitLab 19.6 adds an audit log for AI agents

GitLab 19.6 records what AI agents do through its MCP server. A Ruby flaw is patched, but several NGINX flaws remain open, so admins who run GitLab themselves have settings to enable and a workaround to apply.

Models4 min

Test it with your own documents.

Free for 3 days, then from €24.90 per user per month. Cancellable monthly.

Free for 3 days, then from €24.90 per user per month.Free for 3 days
Try free for 3 days