It does. The EU AI Act regulates AI systems, and an agent built on a model that infers how to respond and act is an AI system. The Act does not have a separate category for agents. What decides the weight of your duties is the agent's use: prohibited, high-risk, subject to transparency duties, or minimal risk.
How does the Act decide which rules apply to an agent?
The Act is risk-based. It sorts AI systems by what they are used for:
- Prohibited practices, banned outright, which have applied since 2 February 2025.
- High-risk systems, including uses listed in Annex III such as creditworthiness assessment and credit scoring of natural persons, and risk assessment and pricing for life and health insurance. These carry the heavy obligations: risk management (Article 9), data governance (Article 10), automatic logs (Article 12), information to deployers (Article 13), human oversight (Article 14) and accuracy, robustness and cybersecurity (Article 15).
- Transparency obligations under Article 50: people must be told when they interact with an AI system, and certain AI-generated content must be marked.
- Minimal risk, where the Act imposes little beyond AI literacy under Article 4.
So a customer support agent is usually in the transparency tier, while the same platform running a credit decision agent is in the high-risk tier. A plain summary of the whole regulation is in what the EU AI Act is.
The Act also applies by role. The provider, who develops the system or has it developed and puts it on the market under its name, carries most of the high-risk design duties. The deployer, who uses it in its business, carries duties under Article 26. If your business is outside the EU, see whether the EU AI Act applies to companies outside the EU.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What is different when the AI system is an agent?
The Act was drafted with systems in mind that produce outputs: a score, a classification, a piece of text. An agent produces actions. It calls tools, changes records and can trigger the next step without a person.
That matters in three places. Classification is harder, because one agent can do several jobs, and one of them may be high-risk while the rest are not. Human oversight under Article 14 has to work at the moment of action, not after, because some agent actions cannot be undone. Logging under Article 12 has to capture what the agent did and why, not only what it output. The argument is developed in the EU AI Act and agents that act.
What may the agent do, and what stays with a person?
The Act does not list permitted agent actions. It requires oversight proportionate to risk. In practice:
| Agent use | Likely tier | What stays with a person |
|---|---|---|
| Answering customer questions | Transparency | Complaints, exceptions, anything outside policy |
| Drafting internal reports | Minimal | Sign-off before use |
| Recommending credit decisions for individuals | High-risk | The decision, and review of any challenge |
| Pricing life or health insurance for individuals | High-risk | Approval of the price and of exceptions |
| Reconciling accounts | Minimal | Postings above the customer's limit |
An illustrative example
Take a bank that deploys one agent with three capabilities: answering card queries, drafting complaint responses, and pre-assessing small personal loans. The bank first treated the agent as one system in the transparency tier. On review, it split the classification by capability. The loan pre-assessment is creditworthiness assessment of natural persons, so that capability is high-risk, and the bank now needs the provider's documentation, assigned human oversight, logs kept under its control and a fundamental rights impact assessment under Article 27. The card queries capability only needs the Article 50 notice.
Which controls meet the Act for agents?
- Classify per capability, not per agent, and record the reasoning.
- Know your role for each agent: provider, deployer, or both.
- Hold actions that a high-risk capability produces until a competent person approves them.
- Log every step the agent takes, with the sources it relied on, not only its final output.
- Tell people they are dealing with an AI system wherever Article 50 applies.
- Train supervisors under Article 4, matched to what they approve.
A practical list for each of these is in the EU AI Act compliance checklist for AI agents, and the cross-regulation view is in AI agent regulation, mapped to controls. For US readers, the contrast with whether the Colorado AI act applies to AI agents is instructive: Colorado looks at consequential decisions, the EU at listed uses.
Questions people ask
Is an agent built on a general-purpose model treated as a general-purpose AI model?
No. The general-purpose model obligations fall on the model provider. Your agent is an AI system built on that model, and it is classified by its own use.
Can one agent be both high-risk and low-risk?
In effect, yes. Classify each capability. If any capability is a high-risk use, the high-risk obligations apply to that use, and it is usually cleaner to separate it.
Does the Act require a human to approve every agent action?
No. It requires oversight appropriate to the risk. For high-risk uses, that means people who can understand, monitor and stop the system, and in practice a hold on decisions with significant effects on individuals.