Cyber, data & AI

Agentic AI

Agentic AI (AI agents) is an umbrella term for AI systems that pursue a goal with little human intervention by planning, executing and adjusting multi-step tasks themselves, often with access to other software, data or systems. There is no legal definition. The EU AI Act treats an AI agent as an AI system as soon as it meets the definition in Article 3(1), without a separate risk category or separate obligations.

Source: Regulation (EU) 2024/1689 (AI Act) Articles 3(1), 14, 26 and 50; Regulation (EU) 2026/1744 Article 1, point 43 (Annex XIV, code AIH 0401); FSB, Sound Practices for Responsible Adoption of AI (consultation, June 2026); COSO, Achieving Effective Internal Control Over Generative AI (2026)

The original text of the EU AI Act (Regulation (EU) 2024/1689) does not use the term. Since the amendment by Regulation (EU) 2026/1744, Annex XIV does mention agentic AI, as an example under the code for emerging AI technologies that is used to set the scope of the designation of notified bodies. No definition follows from this. Whether an agent is prohibited, high-risk or subject to transparency obligations depends, as for any AI system, on its use. In a 2026 publication on internal control over generative AI, COSO describes agents as models that plan, execute and adapt multi-step tasks to achieve a defined goal, often by interacting with external tools, data or environments.

The risk lies in the combination of autonomy and permissions. In a consultation report of June 2026, the Financial Stability Board (FSB) warns that real-time human oversight of agents becomes impractical at scale, so that agents may pursue objectives or take actions that deviate from the institution's intentions or risk appetite without staff being aware or able to intervene in time. The US banking supervisors explicitly exclude generative and agentic AI from their revised model risk guidance SR 26-2 (April 2026), because these technologies are novel and evolving rapidly.

For the management board and the supervisory board, the issue is one of boundaries: which systems and data an agent may use, which actions require human approval, how actions are logged and who the owner is. For high-risk systems, the AI Act requires, from 2 December 2027 (Annex III), human oversight by people with the necessary competence, training and authority (Articles 14 and 26). The difference from generative AI is that an agent does not only create content but also acts; the difference from traditional process automation is that the steps have not been laid down in advance by people.

Last updated: · Editor: ONE Risk Advisory

Frequently asked questions

Does agentic AI fall under the EU AI Act?

Yes, as soon as the agent meets the definition of an AI system in Article 3(1). Which obligations apply depends on the use: an agent that assesses the creditworthiness of natural persons falls under the high-risk rules of Annex III from 2 December 2027. The Act has no separate regime for agents; Annex XIV mentions agentic AI only in the context of designating notified bodies.

What is the difference between an AI agent and a chatbot?

A chatbot answers questions and generates text. An agent also takes actions itself, such as retrieving data, running reconciliations, assigning tasks or escalating exceptions, and adjusts its approach along the way. The risk therefore shifts from incorrect output to incorrect actions. Systems that interact directly with people must make clear that people are dealing with AI (Article 50(1) of the AI Act).

Who is responsible when an AI agent makes a mistake?

The organisation that deploys the agent; buying it from a supplier does not change that. For insurers, EIOPA (2025) states that undertakings are ultimately responsible for the AI systems they use, whether developed in-house or with third-party providers. An organisation therefore records for each agent an owner, the permitted actions and the points at which human approval is required.

← All terms