Under the EU AI Act, an AI agent is generally treated as an AI system, not as a separate legal category. Which rules apply depends on the system’s intended purpose, its risk classification and the roles of the organizations involved. That means the label “agent” alone does not determine an organization’s obligations.
The European Commission says the term “AI agent” has no legal definition and is used for different kinds of systems. Its agent-specific interpretation is preliminary. The Commission’s FAQ on AI agents says agents typically qualify as AI systems and may also involve a general-purpose AI (GPAI) model.
Which rules apply, and when?
Several AI Act obligations were already in application by October 28, 2026. Article 50 transparency rules began applying on August 2, 2026, while the Act’s high-risk requirements have later dates that depend on the system’s category.
| Date | Milestone |
| February 2, 2025 | Prohibitions on specified AI practices and AI-literacy provisions began applying. |
| August 2, 2025 | Obligations for GPAI models began applying. |
| July 27, 2026 | The Digital Omnibus entered into force and changed the high-risk application timetable. |
| August 2, 2026 | The Act became generally applicable, including Article 50 transparency rules for covered systems. |
| December 2, 2027 | High-risk rules for systems in Annex III are scheduled to apply. |
| August 2, 2028 | High-risk rules for AI embedded in products covered by Annex I are scheduled to apply. |
Article 50 covers distinct transparency duties. A provider must inform people when a system is intended to interact directly with them, unless the AI nature of the interaction is obvious in context. Providers of generative systems also have machine-readable marking duties. Deployers have separate disclosure duties for specified deepfakes and certain public-interest text. A voice bot designed to interact directly with people, for instance, falls under the interaction-disclosure rule unless that interaction is obvious.
For systems already on the market, the transition period for Article 50(2)’s marking and detection duties runs until December 2, 2026. The other dates in the table concern different obligations, so the Article 50 timetable does not make every agent a high-risk system.
Risk classification sets the scope of high-risk duties
An AI system is high-risk when its intended purpose and the statutory criteria place it in a high-risk category. The Commission identifies examples including certain uses in recruitment, education, creditworthiness and critical infrastructure. An agent’s autonomy or access to tools does not, by itself, make its system high-risk.
| Rule area | Relevant trigger or assessment | Who or what it concerns |
| Article 50 transparency | Intended interaction with people, generation of content, or specified disclosure situations | Providers and deployers have different duties, depending on the system and content. |
| High-risk requirements | Intended purpose falls within the Act’s high-risk criteria | The system and the providers and deployers responsible for it. |
| Article 25 provider status | Specified branding, substantial modification, or a change of intended purpose that makes a system high-risk | A downstream organization may take on the provider role. |
| GPAI systemic-risk assessment | Model-level assessment that considers factors including autonomy and tool access | The underlying GPAI model and its provider, not automatically every agent using it. |
For high-risk systems, providers have duties that include risk management, data governance, technical documentation, logging, conformity assessment and post-market monitoring. Deployers must use the system in accordance with its instructions, assign competent and authorized human oversight, monitor its operation and keep logs under their control, among other requirements.
The Act also requires effective human oversight for high-risk systems. That includes the ability to monitor a system, disregard or override its output, and intervene or stop operation when appropriate. It does not create a blanket requirement for a person to approve every action taken by every AI agent.
When can a downstream organization become the provider?
Article 25 can shift provider responsibilities to a distributor, importer, deployer or another third party in specified circumstances. These include putting the organization’s name or trademark on a high-risk system, substantially modifying a high-risk system, or changing a system’s intended purpose so that it becomes high-risk. An ordinary customization does not automatically trigger a change of provider role.
That distinction matters when an organization builds an agent using another provider’s system and then changes how it is presented or used. The relevant question is whether one of Article 25’s conditions applies, not simply whether the organization adjusted the system.
The model behind an agent gets a separate assessment
A GPAI model and an AI agent that uses it are related but distinct subjects under the Act. Autonomy and tool access are factors in assessing whether an underlying GPAI model presents systemic risk. A model-level designation can bring additional duties for its provider, including assessment, risk mitigation, incident reporting and cybersecurity measures. Tool use does not automatically make every agent or underlying model systemic-risk AI.
AI Act duties do not decide compensation by themselves
The AI Act assigns obligations to providers and deployers and preserves relevant rights and remedies under other EU and national laws, including remedies concerning damages. Its compliance duties alone do not determine who must compensate someone harmed in a particular case; that question depends on the applicable law and facts. The text of Regulation (EU) 2024/1689 sets out the statutory roles and obligations.
