In most cases, yes. The AI Act does not define an "AI agent" or treat it as a separate category: according to the Commission's AI Act Service Desk FAQ, an agent will usually constitute an AI system, typically built around a general-purpose AI model. The rules for AI systems (transparency, prohibited practices, high-risk depending on use) therefore apply to your agent.
What exactly does the Commission say?
The Service Desk FAQ makes several points:
- "AI agent" has no legal definition, and agents are not a separate category. The existing definitions of an AI system (Article 3(1)) and a general-purpose AI model (Article 3(63)) are enough to cover them;
- an agent usually combines a general-purpose AI model with some form of interface, which counts as a system component (recital 97), for example to issue function calls;
- Article 50 applies from 2 August 2026 if the agent is meant to interact with people or generate content;
- if the agent is classified as high-risk, Chapter III applies, from 2 December 2027 or 2 August 2028.
The Commission describes these views as preliminary, and the AI Office is monitoring how agents develop. They are still the best official guidance available today.
Should you assess the agent or the whole workflow?
The whole workflow. Article 3(1) defines an AI system as a machine-based system designed to operate with varying levels of autonomy that infers from its inputs how to generate outputs (predictions, content, recommendations or decisions) that can influence physical or virtual environments. The Commission's guidelines on the AI system definition unpack these criteria.
An agent built in n8n or Make is more than the model call. The trigger, system prompt, available tools (sending email, writing to a CRM, querying a database, making payments) and output together make up the system. Two consequences:
- intended purpose is read at workflow level: the same model carries no particular risk in an agent that summarises meetings, and is high-risk in an agent that shortlists candidates (Annex III, point 4) or assesses a customer's creditworthiness (point 5(b));
- tools widen the reach of the outputs: an agent that acts (sends, edits, decides) directly influences its environment, which matters for risk analysis, human oversight and logging.
Whoever assembles the agent and puts it into service under its own name is in principle its provider; how that plays out between an agency and its client is covered in automation agencies on n8n and Make: provider or deployer?.
What transparency rules apply to an agent?
The same as for any AI system, depending on what the agent does (Article 50):
- if it talks directly with a person (chat, conversational email, phone call), the provider must make sure that person knows they are interacting with an AI, at the latest at the first interaction (paragraphs 1 and 5);
- if it generates text, images, audio or video, the provider must mark that content in a machine-readable way (paragraph 2);
- if the deployer publishes deepfakes or public-interest text generated by the agent, the deployer must disclose it (paragraph 4).
The Commission's Article 50 Q&A expressly lists AI agents among the systems covered, where the AI itself communicates with the person rather than through a human intermediary. An agent that writes to or calls third parties on a company's behalf must therefore present itself as an AI; also saying on whose behalf it acts is good practice, though not a written requirement of Article 50. Chatbots are covered in detail in is my chatbot covered?.
What guardrails against prohibited practices?
The FAQ highlights two Article 5 prohibitions, in force since 2 February 2025, that may call for safeguards from the design stage:
- subliminal, manipulative or deceptive techniques that materially distort a person's behaviour and cause, or are likely to cause, significant harm (paragraph 1(a));
- exploiting vulnerabilities due to age, disability or a specific social or economic situation (point (b)).
A sales agent tuned to "close at any cost" or a debt-collection agent that targets people in financial difficulty are examples of risks to address. From 2 December 2026, the Omnibus also bans systems that generate non-consensual intimate images of an identifiable person or child sexual abuse material (points (ba) and (bb)), which directly concerns agents with image-generation tools. The guidelines on prohibited AI practices explain these concepts.
For model providers, the FAQ also notes that a model's level of autonomy or tool use can be decisive in classifying it as a general-purpose AI model with systemic risk (Article 51(1)(b) and Annex XIII). That concerns model developers, not the team integrating an agent.
What should you do in practice?
- Declare the agent as an AI system, with its real purpose and the list of its tools.
- Check whether that purpose falls under Annex III.
- Keep tools to the minimum needed and require human approval before irreversible actions.
- Show the Article 50 disclosure on every channel where the agent talks to people.
- Log decisions and tool calls, and keep a record of the controls (see how to prove compliance).
VerifactAI classifies the agent as a system in its own right, identifies the resulting obligations, can block or warn in n8n or Make before the model call, and records each decision with its evidence.
Sources
- AI Act Service Desk FAQ: how are AI agents addressed within the AI Act?
- Regulation (EU) 2024/1689, EUR-Lex
- Guidelines on the definition of an AI system
- Guidelines on prohibited AI practices
- Guidelines on transparency obligations
- Commission Q&A on Article 50
This article is general information, not legal advice.