AI agents in business: what the AI Act and GDPR require

Businesses are piloting AI agents: software that does not just suggest tasks but carries them out, sometimes without a specific instruction from a person. Management and IT rightly ask whether the AI Act or the GDPR demand anything special for that.

The short answer: not for the agent as such. “AI agent” is not a term in the AI Act. The existing rules apply, depending on the agent’s purpose and your role. Here we call agents “proactive” when they act, for example, in response to an event. That is a working definition, not a legal term.

This article shows which rules these are, where agents make them harder to apply and what to clarify before going live.

No. The AI Act knows neither “AI agent” nor “agentic AI” as a category with its own obligations. An agent is an AI system within the meaning of Art. 3(1), that is, a system designed to operate with varying levels of autonomy. That is still true after the Digital Omnibus (Regulation (EU) 2026/1744).

The same rules therefore apply to agents as to any AI system, depending on intended purpose and role: the prohibitions in Art. 5, classification as a high-risk system under Art. 6 and Annexes I and III, the transparency duties in Art. 50 and the duty to take AI literacy measures under Art. 4. If the agent is built on a general-purpose AI model, the additional obligations for that model fall on the provider of the model.

Your role determines which obligations land with you. If you use a ready-made agent, you are a deployer. That also applies to tools from providers in third countries: a German company using a US tool for work is a deployer. Deployers are subject to Art. 4 and, depending on the use, Art. 50 and Art. 26.

You become a provider faster than many assume. The table shows the most common cases.

An example from the mid-market: a machine builder with €12 million in revenue uses an agent delivered ready-made by a software house. It is a deployer. If it has the same agent developed for its own sales and puts it into service under its own name, it is provider and deployer at once.

More on these duties is on our page about deployer obligations under the AI Act.

High-risk status arises in only two ways: as a safety component or product under Annex I with third-party conformity assessment, or as a use case under Annex III (Art. 6). General features do not make an agent high-risk, such as acting independently, making decisions about people or seeing sensitive data. Those are rather checkpoints under the GDPR (Art. 22, Art. 35).

The obligations for high-risk systems under Annex III apply from 2 December 2027. Deployers must then, among other things, assign human oversight to competent, authorised persons, monitor operation and keep automatically generated logs for at least six months (Art. 26(2), (5) and (6)).

The Commission intends to adopt the final classification guidelines (Art. 6(5)) by the end of 2026. So far there is only a draft of 19 May 2026 (as of 26 October 2026). The final version may differ.

That depends on who acts. If a person drafts an email with AI and sends it themselves, that triggers no labelling duty under Art. 50. If, on the other hand, an AI system answers emails on its own, that is, in direct interaction with people, the provider must design it so that recipients are informed, unless that is obvious (Art. 50(1)).

A note in the email signature is not required by law. The classification of AI agents follows the Commission’s guidelines on Art. 50 of 20 July 2026, which are not legally binding. Art. 50 has applied since 2 August 2026.

Infringements of Art. 50 can be fined up to €15 million or 3 % of worldwide annual turnover (Art. 99(4)). For an SME (fewer than 250 staff) with €12 million in worldwide annual turnover the cap is 3 %, that is €360,000 (Art. 99(4) and (6)). The actual amount depends on the individual case (Art. 99(7)).

Details on labelling are on our page about Art. 50 of the AI Act.

The GDPR has no special rules for agents either. Provider and deployer under the AI Act are different roles from controller and processor under the GDPR, and the GDPR remains unaffected (Art. 2(7) AI Act). With agents, three questions are especially sharp.

Logs you keep about the agent often contain personal data. For that you need a basis under the GDPR. So record what you really need and no more. That is a recommendation, not a legal duty.

Not everything that agents raise is settled. The AI Act creates no claims for damages of its own and no criminal offences. Claims follow national law, for example tort law (§ 823 German Civil Code) and, for personal data, Art. 82 GDPR.

In its note (exploratory note with CIANum, 20 July 2026, non-binding), the CNIL takes the view that agents make it harder to attribute responsibility. A controller under the GDPR must nevertheless remain identifiable. The AI Act contains no express liability regime. The Dutch Autoriteit Persoonsgegevens warned on 12 February 2026 against using AI agents such as OpenClaw on systems with sensitive or confidential data, and stated that organisations and users remain responsible for GDPR compliance. It also asks for clarification at EU level that autonomous AI agents fall under the AI Act. That is a demand by the authority: in law, agents are already AI systems under Art. 3(1).

The steps are a recommendation. The AI Act contains neither a four-eyes principle nor a general approval duty before an agent is used. Only human oversight for high-risk systems is regulated.

An agent is not a special case in law. It is an AI system, and what counts is what it is intended for and what role you have in that.

The classification can only be evidenced if you have recorded for each agent what it may do, who is responsible and what it has done. That is the basis for every later review, whether by an authority or by customers.

This article is for general information only and does not constitute legal advice. Last updated: 26 October 2026. It relies in one point (classification guidelines under Art. 6(5)) on the Commission draft of 19 May 2026; the final version may differ.