A model is a trained or configured computational artefact that infers outputs from inputs. A component is a part of a product or system, which may or may not itself be an AI system. An AI system is the machine-based system with an intended purpose that is placed on the market or put into service. Classification under Articles 5 and 6 is applied to the AI system in its intended use, not to a model in isolation, so the three layers must be named separately before any risk conclusion is drawn.
Why do teams confuse these three layers?
Everyday language calls the whole product the model, the model the system, and the system the feature. In product discussions that shorthand is harmless, but in a regulatory record it hides the object being assessed. A vendor may sell a model, a platform may host it, an integrator may embed it, and a customer may deploy the resulting system. Each step adds purpose, context and controls. If the record names only the vendor's product name, a reviewer cannot tell whether the classification considered the model alone, the component in a host product, or the deployed system with its workflow. Naming the layers is therefore the first control, not a matter of presentation.
What counts as a model in this separation?
A model is the artefact that performs inference: weights, parameters, configuration and the interface through which it receives inputs and returns outputs. It may be a general-purpose AI model, a narrow task model, or a fine-tuned variant. Its capabilities are described in general terms, such as the tasks it can perform and the data it was trained on, but it has no deployment context of its own. Two systems built on the same model can behave differently because of prompts, tools, data sources and human review. The model record should therefore capture the model name, version, supplier, licence or supply terms, and the information the supplier provided about capabilities and limitations.
For the broader regulatory framework, see EU AI Act fundamentals.
What counts as a component?
A component is a part of a larger product or system. It may be a model, a software library, a sensor, a user interface element, or a safety function. Article 3 defines a safety component as a component of a product or of an AI system which fulfils a safety function, or whose failure or malfunctioning endangers the health and safety of persons or property, with the Omnibus amendments adding detail on non-safety uses. A component analysis asks what the component does, what happens if it fails, and whether its function is safety-related. A component that is not itself an AI system can still matter for classification when it forms part of a high-risk product, so components should be inventoried even when they are not the headline feature.
What counts as the AI system?
The AI system is the complete machine-based system as supplied and used for its intended purpose, including the model, the surrounding software, the user interface, the data flows, the human oversight arrangements and the operating instructions. It is the system, in that complete form, that is screened against Article 5, classified under Article 6, and assessed for Article 50 duties. A change to any material part of that complete form, such as a new data source, a new user group, or a change in how outputs are acted on, can require the classification to be revisited. The system record should therefore describe the whole deployed configuration, not only the model at its centre.
How should the layers be recorded together?
A useful record uses three linked entries. The model entry names the model and its supplier information. The component entry names each part, its function, and its failure effect. The system entry names the intended purpose, the users, the affected people, the deployment context, and the classification reasoning. Links between the entries show which model version sits in which component and which system version. When a supplier updates a model, the link makes it visible which systems are affected. When a deployer changes a workflow, the system entry changes while the model entry may stay the same. That separation prevents a model update from silently changing a system classification, and vice versa.
Continue with what counts as an AI system.
Where does this separation stop helping?
Separating layers clarifies the object of assessment, but it does not classify anything by itself. A team still needs to test the complete system against Article 5 and Article 6, check Article 50, and assign roles under Article 3 and Article 25. This guide does not decide whether any particular model, component or product is an AI system, a safety component, or a high-risk system. Those conclusions need the intended purpose, the failure analysis, and the deployment facts, confirmed with a competent authority or legal adviser where the boundary is contested or the consequences of error are high.
Frequently asked questions
Is a software library an AI system?
Not necessarily. A library may be a component. Whether the wider system that uses it is an AI system depends on the complete system, its inference mechanism, its intended purpose and its outputs, assessed as a whole.
Does updating a model always change the system classification?
No, not always, but every material model update should trigger a review. The review checks whether the intended purpose, the inference behaviour, the user group or the failure effects have changed enough to affect the earlier classification.
What is a safety component in simple terms?
It is a component that fulfils a safety function for a product or an AI system, or whose failure or malfunctioning endangers health and safety of persons or property. The current consolidated definition and the Omnibus additions on non-safety uses should be checked in the legal text before relying on the label.
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 ↗