An AI system under Article 3(1) is a machine-based system designed to operate with varying levels of autonomy, that may exhibit adaptiveness after deployment, and that infers from input how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments, for explicit or implicit objectives. Software that only executes fixed rules written by a human, without inference, does not meet that definition on its face, but the boundary is fact sensitive and must be tested against the actual behaviour of the system, not its marketing label.
Why is the threshold question so important?
Every later question in the Act presupposes an AI system or a general-purpose AI model. Prohibited practices, high-risk classification, transparency duties and actor obligations all attach to those defined objects. If the technology is not an AI system, the AI Act analysis stops, although other laws may still apply. If it is an AI system, the analysis continues through intended purpose, role, classification and timing. Teams that skip the threshold question often spend effort on the wrong risk tier, or miss a duty entirely because they treated a system as ordinary software. A short threshold record, made before classification, saves rework and makes later review possible.
What does machine-based and inference mean here?
Machine-based means the system runs on machines rather than as a purely human process. Inference means the system derives how to generate its output from the input it receives, for explicit or implicit objectives, rather than merely applying a fixed instruction that fully determines the output in advance. Machine learning models, including large language models and many statistical models, will commonly involve inference. A lookup table, a fixed decision tree written entirely by hand, or a spreadsheet formula that applies the same rule to every input does not obviously involve inference in this sense. The question is about how the output is produced, not about whether the software is sophisticated or commercially described as AI.
For the broader regulatory framework, see EU AI Act fundamentals.
What role do autonomy and adaptiveness play?
The definition requires design for operation with varying levels of autonomy, and notes that the system may exhibit adaptiveness after deployment. Autonomy here means the system can operate without continuous human direction for the steps it performs, even if a human sets the objective, reviews the output or can intervene. Adaptiveness is only a possibility, signalled by the word may, so a system that does not learn after deployment can still be an AI system. Conversely, adaptiveness alone does not make a system an AI system if the inference element is absent. Both features should be described in the record in concrete terms: who directs each step, what the system decides itself, and whether its behaviour changes after deployment.
How should a team test a real product against the definition?
Start by describing what the system receives, what it produces, and how the output is generated. Identify the model or inference mechanism, the training or configuration that shapes it, and the degree of human direction at run time. Then ask whether the output can influence a physical or virtual environment, including by informing a human decision that changes what happens next. Record the evidence examined, such as technical documentation, model cards, supplier descriptions and observed behaviour, and note any supplier claim that was not verified. If the mechanism is unclear, treat the threshold question as open and seek further technical evidence before concluding that the Act does not apply.
What are common borderline cases?
Borderline cases include statistical tools that blend fixed rules with learned components, search and ranking features that infer relevance, optimisation tools that adapt to new data, and human-in-the-loop workflows where the machine proposes and the human disposes. In each case, the label on the product is not decisive. A hybrid system may contain an AI system as a component even if the wider product is not itself an AI system, and a feature that only formats or displays human-written content may not infer anything. The safest record separates the component from the product, states which part performs inference, and explains why the boundary was drawn where it was.
Continue with Article 3 definitions.
What does this guide leave undecided?
This guide explains the test in Article 3(1) and how to evidence it. It does not decide whether any named product, feature or model is an AI system. That determination depends on the actual inference mechanism, the autonomy of its operation, and the influence of its outputs, and it can change when a supplier updates a model or a team changes how a feature is used. Where the mechanism is proprietary or the facts are incomplete, the threshold question should stay open in the record, with an owner and a date for closing it, and confirmation should be sought from a competent authority or legal adviser rather than assumed from a vendor label.
Frequently asked questions
Does using the word AI in marketing make software an AI system?
No. The legal test is the Article 3(1) definition, applied to how the system actually operates. Marketing language is evidence of how a system is presented, but it does not replace the inference, autonomy and output analysis.
Does a human reviewer mean the system is not autonomous?
Not necessarily. The definition allows varying levels of autonomy, and a system can be an AI system even where a human reviews or can override its output. What matters is whether the system itself infers how to generate outputs that can influence an environment.
Can part of a product be an AI system while the whole product is not?
Yes, in practical terms. A product can contain an AI component that meets the definition, and the analysis should identify that component, its intended purpose and its outputs separately from the wider product, before deciding what follows for the product as a whole.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689 ↗AI Act Service Desk timeline ↗European Commission regulatory framework for AI ↗