Glossary · AI governance
The label AI system depends on how software reaches its outputs.
Under the EU AI Act, a machine-based system that infers from its inputs how to produce outputs such as predictions, content, recommendations or decisions, with some autonomy.
Back to the glossary01 · In plain words
In plain words
Under the EU AI Act, a machine-based system that infers from its inputs how to produce outputs such as predictions, content, recommendations or decisions, with some autonomy.
Calling a product AI does not make it one in law, and leaving the word off does not take it out. The EU Artificial Intelligence Act (the AI Act) looks at inference: does the software work out from its inputs how to produce an output, rather than only follow rules a person wrote? Add some autonomy and outputs that can affect a physical or digital setting, and you are probably looking at an AI system.
Spelled out: Artificial intelligence system, as defined in Article 3(1) of the EU AI Act.
02 · Why it matters
Why it matters
Who cares about this, and what depends on it.
Every duty in the AI Act hangs off this definition. If a tool is not an AI system, the risk tiers, the transparency rules and the provider and deployer duties do not reach it. If it is, the next questions follow fast: is it prohibited, high-risk, or subject only to disclosure? European buyers now ask vendors to state which features count, and an ISO/IEC 42001 auditor expects an AI inventory that applies the same test every time. A wrong call costs you either way. Over-claiming buries a small team in paperwork. Under-claiming leaves a gap that a regulator or a customer's reviewer will find.
03 · Where you meet it
Where you will meet it
The documents, meetings and reports where it appears.
- Article 3(1) of the AI Act
- The definition itself: machine-based, some autonomy, possible adaptiveness after deployment, objectives, inference, and outputs that influence an environment.
- The Commission's guidelines on the definition
- Guidance issued in 2025 to help sort borderline software, such as simple rule engines and basic statistical tools.
- Your AI inventory
- The register an AIMS under ISO/IEC 42001 keeps, with each tool marked in or out and the reason written down.
- Due-diligence forms from EU customers
- Questions asking which features use AI, and in what role you supply each one.
04 · An example
A worked example
An illustration, not a real client.
Illustrative: two ranking features in one HR product
A Pune HR-tech company sells a candidate tracker with two ways to sort applicants. The first sorts by rules a recruiter sets: years of experience, notice period, a required certificate. The second uses a model trained on past hiring outcomes to score each résumé for fit. Marketing had called both 'smart ranking'. Under the inference test, the rule sort is ordinary software. The trained scorer is an AI system, and because it filters job applicants it sits in an employment use area listed in Annex III, so it is likely high-risk. The inventory now records the two features separately, with the reasoning, and only the second goes into risk and impact work.
05 · Easy to mix up
Easy to mix up with
Neighbouring terms that mean something different.
- AI model
- A model is a component, such as a trained network. An AI system wraps one or more models with inputs, an interface and a purpose. The Act treats general-purpose models in a chapter of their own.
- Automation
- A script that follows fixed rules written by people usually lacks the inference element, however sophisticated it looks.
- Agentic AI
- An agent is one kind of AI system that plans and acts with more independence. Every agent is an AI system; most AI systems are not agents.
06 · What to do
What to do with it
Practical steps, and the guides that go deeper.
- List every feature and internal tool that might use a trained model, including ones bought from vendors.
- Apply the Article 3(1) elements to each and write down why it is in or out.
- For each one that is in, record your role and a first view of its risk tier.
- Recheck a feature whenever it moves from fixed rules to a learned model.
- GuideISO 42001 and the EU AI ActLaw vs standard, the risk tiers, dates, and the article-by-article map.
- GuideAI governance & the AIMSPrinciples, clauses 4–10, the operating model and the first five moves.
Sources: Regulation (EU) 2024/1689 (the AI Act), Articles 2 and 3 and Annex III, from EUR-Lex; European Commission guidelines on the definition of an AI system; ISO/IEC 22989 for general AI vocabulary.
Questions about AI system
Is a chatbot built on someone else's model an AI system?
Usually yes, and you may be its provider.
The product you build around a model, with its prompts, interface and purpose, is normally an AI system in its own right. Whoever places it on the market under their own name is generally its provider, even though the model underneath came from elsewhere.
Does a statistical forecast count?
Not always. It turns on inference.
A tool that applies a fixed formula may fall outside. Where a forecasting tool learns patterns from data and its outputs steer decisions, treat it as in scope until you have a reasoned view otherwise, and keep that reasoning in the inventory.
Does the definition matter for a company outside the EU?
Yes, if the system reaches the EU.
Article 2 covers providers placing AI systems on the EU market wherever they are based, and providers and deployers outside the EU whose system's output is used in the Union. An Indian company selling into Europe should run the test.
Want this applied to your own systems?
Our team can walk through where this term touches your scope, your contracts and your evidence.