An AI agent built on a language model, with tools and a written brief, almost always meets the EU AI Act's definition of an AI system. The definition turns on inferring from inputs how to generate outputs that influence an environment, which is what an agent does. Every other duty under the Act depends on this answer.
What does the Act's definition say?
The Act defines an AI system, in substance, as a machine-based system that:
- is designed to operate with varying levels of autonomy;
- may exhibit adaptiveness after deployment;
- for explicit or implicit objectives, infers from the input it receives how to generate outputs;
- produces outputs such as predictions, content, recommendations or decisions;
- whose outputs can influence physical or virtual environments.
The key element is inference. Systems built only on rules defined by people to execute operations automatically are not AI systems in the Act's sense. A spreadsheet formula, a fixed workflow or a decision table written by an analyst falls outside. A system that learns, reasons or derives its outputs from patterns rather than from rules a person wrote falls inside.
Some uses are excluded from the Act altogether, such as AI used only for scientific research and development, purely personal non-professional use, and military and national security purposes. None of these will usually apply to a business deploying agents. For a primer on the law, see what the EU AI Act is.
Why does this question come before every other?
Because nothing else in the Act applies until the answer is yes. Risk classification, prohibited practices, transparency to people, AI literacy, deployer duties, logs and oversight all attach to AI systems. A firm that concludes a tool is not an AI system has, in effect, decided that the Act does not apply to it at all.
That makes the conclusion high-stakes, and it should be written down. An auditor's first question about any automated tool that influences customers may be whether it is an AI system, and "we assumed not" will not do. The question whether an AI agent is a high-risk AI system can only be asked once this one is answered.
How does the definition apply to an agent?
Take an agent's usual parts:
- A model that interprets instructions and content and generates text, plans and tool calls. This is the inference engine.
- A brief that tells it its role, scope and rules.
- Tools and connections that let its outputs act on systems: send an email, update a record, initiate a payment.
- Sources it reads at runtime.
Each element of the definition is present. The agent operates with some autonomy, because it decides which steps to take. It infers from input how to generate outputs, because the model derives its plan and wording from the content it reads. Its outputs, including tool calls, influence virtual environments directly. Adaptiveness is optional in the definition, so an agent that does not learn after deployment still qualifies.
The brief does not change the answer. Writing detailed rules for an agent does not turn it into a rule-based system, because the model still interprets those rules and decides how to apply them.
Where are the edges?
Hybrid systems. A workflow that is mostly fixed rules but calls a model for one step, such as classifying an email, contains an AI system. Treat the model-driven step, and anything that acts on its output, as in scope.
Deterministic tools an agent calls. A payment interface or a rules engine is not an AI system by itself. When an agent decides to call it and with what parameters, the decision is the AI system's output.
What is the system boundary? A platform may run many agents. For classification, we think each agent, and often each capability, should be treated as the unit, because the Act classifies by intended purpose and purposes differ between agents. The provider of the platform may treat the platform as its system; the deployer should classify what it actually uses.
Simple automation dressed as AI. Some products marketed as agents are scripted. If no inference is involved, the Act may not apply. Check what the product does, not what it is called.
How do you record the answer?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Decide whether each tool is an AI system | Screening question in the procurement and change process | Screening record per tool |
| Identify the unit of classification | Register lists each agent and capability as a separate entry | Register structure |
| Document hybrid systems | Map showing which steps use inference and which are fixed rules | System diagram with AI steps marked |
| Revisit on change | New model calls or tool connections trigger re-screening | Change records with screening outcomes |
| Record who decided | Named owner signs off each conclusion | Signed screening records |
An AI agent register for the EU AI Act is the natural home for these entries, with a column for the AI system decision ahead of the tier.
What will a supervisor or auditor ask?
- Which of your automated tools did you decide are AI systems, and which are not?
- Show us the reasoning for any you decided are not.
- Where a workflow mixes rules and a model, which parts did you treat as in scope?
- Who made these decisions?
Enforcement sits mainly with national authorities and, for financial firms, often with sector supervisors; see who enforces the EU AI Act for AI agents.
What does the definition not settle?
It does not decide your role. Whether you are provider or deployer is a separate question; see what a provider is under the EU AI Act. It does not decide the risk tier. And it does not decide whether other law applies: GDPR, conduct rules and sector rules govern automated tools whether or not they are AI systems. For the full map, see AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Is a chatbot with scripted answers an AI system?
If it only follows scripts and decision trees written by people, it may not be. If it uses a model to interpret questions or generate answers, it is. Many products combine both, so check the parts.
Does an agent that only reads and drafts count as an AI system?
Yes. Generating drafts, summaries or recommendations is an output under the definition, and influencing a person's decision is influencing an environment. Its risk tier may be minimal, but it is in scope.
If we change the model behind an agent, is it a new AI system?
It remains an AI system either way. Whether the change is a substantial modification depends on its effect on compliance and intended purpose, and should be assessed through change control.