The EU AI Act and agents that act

The EU AI Act regulates AI systems by what they are used for, not how they work. An agent that acts is still an AI system, and its duties depend on its uses: some minimal risk, some carrying transparency duties, some high-risk. Because one agent can serve several uses, classify each capability, not the agent as a whole.

How does the Act's risk-based structure work?

The EU AI Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024, and its obligations phase in over several years. Prohibitions applied from 2 February 2025, and obligations for general-purpose AI models from 2 August 2025. Application dates for high-risk obligations have been subject to proposed changes, so check the current position before planning around them. Our explainer on what the EU AI Act is gives the overview.

The Act sorts uses into four groups. Some practices are prohibited. Some uses are high-risk, including those listed in Annex III, such as creditworthiness assessment and credit scoring of individuals, and risk assessment and pricing for life and health insurance. Some uses carry transparency obligations, such as telling people they are interacting with an AI system under Article 50. Everything else is minimal risk.

What changes when the AI system takes actions?

The Act was drafted with systems in mind that produce outputs, such as predictions, recommendations or content, which then influence decisions. Agents go further: they carry out the action. That does not create a new category, but it sharpens several existing duties.

Human oversight. Article 14 requires high-risk systems to be designed so people can oversee them effectively, including understanding outputs, deciding not to use them and intervening or stopping the system. For a system that only recommends, the person decides anyway. For an agent that acts, oversight has to be built into the action path: held actions for named people, the ability to lower authority or stop the agent at once.

Accuracy, robustness and cybersecurity. Article 15 expects high-risk systems to perform consistently and resist attempts to alter their behaviour. For an agent, the most direct threat is manipulated input, such as instructions planted in a document it reads, leading it to take an action nobody intended. Controls that check each action against the agent's bounds, regardless of its reasoning, address this directly.

Record-keeping. Article 12 requires automatic logging for high-risk systems. For an agent, the log needs to make each action traceable, which in practice means recording what was decided and why, not just what happened.

Why should agents be classified per capability?

Because one agent can span several risk categories. A customer service agent at a bank might answer product questions, help with complaints and, if allowed, assess whether a customer qualifies for a credit limit increase. The first use is minimal risk with a transparency duty. The last falls within creditworthiness assessment, a high-risk use.

Classifying the agent as a whole would either over-regulate the simple uses or under-regulate the high-risk one. Classifying per capability lets each carry the duties that fit it, and lets the firm decide whether to keep high-risk capabilities out of an agent altogether. Our guide to building an AI agent register for the EU AI Act shows how to record this.

Who carries the obligations?

The Act distinguishes providers, who develop a system or have it developed and place it on the market or put it into service under their name, from deployers, who use a system under their authority. Each has different duties. Most firms using an agent platform are deployers; our explainer on what a deployer is under the EU AI Act covers that role. Deployers of high-risk systems must, among other things, use them according to instructions, assign human oversight to competent people, monitor operation and keep logs under their control, under Article 26. Some deployers must carry out a fundamental rights impact assessment under Article 27.

Firms that build their own agents may be providers too. The boundary matters and depends on facts such as who develops the system and under whose name it is used.

What does this look like in customer service?

Customer service is where many firms deploy their first agent, and most of it sits in lower-risk territory, with Article 50's duty to tell people they are interacting with an AI system as the main specific obligation. The risk rises when service agents start making decisions about access to products or pricing. Our page on the EU AI Act for AI agents in customer service explores where the lines fall.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

What is the trade-off?

Per-capability classification takes more effort than a single label per agent. It needs an inventory of what each agent does, a judgement on each use and a willingness to split agents or remove capabilities where a high-risk use would bring disproportionate duties.

The single label is simpler and usually wrong. List every capability of every agent, classify each against the Act's categories, and attach the duties that follow. Our EU AI Act compliance checklist for AI agents gives a starting structure, and the wider mapping of rules to controls is in our guide to AI agent regulation, mapped to controls.