The term artificial intelligence system (or AI system) is defined in Article 3(1) of the European Artificial Intelligence Regulation (AI Act).
This concept is important because only systems that meet the definition of an AI system within the meaning of that article are subject to the Regulation’s requirements.
Given the rapid pace of technological change, it is important to note that this definition remains flexible and should not be applied mechanically: each system must be assessed on a case-by-case basis, according to its own characteristics.
According to Article 3(1) of the Artificial Intelligence Act (AIA), an artificial intelligence system (AI system or AIS) is defined as follows: a machine-based system designed to operate with varying levels of autonomy, that may adapt after deployment, and that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments.
So, an AI system is described as a machine-based system, designed to operate with varying levels of autonomy (ranging from fully human-controlled systems to those capable of making decisions independently), and capable of adaptation after deployment (for example by learning from new data or adjusting its behavior).
Criteria for defining an AI system within the meaning of the AI Act
According to the European Commission guidelines, which supplement this definition, a system falls within the AI definition under the AI Act if, at some point in its lifecycle (design, development or use), and therefore not necessarily at every stage, it displays some of the following characteristic elements:
1. Automated machine-based system
The AI system relies on hardware and software infrastructure (processing units, memory, computer programs) and operates in an automated manner. Systems based on emerging technologies (e.g. quantum computing, biological or organic systems) are also included, provided they have processing capabilities.
2. Variable levels of autonomy
An AI system can operate with a more or less high degree of autonomy, meaning with partial independence from human intervention. The definition excludes systems that are entirely manually controlled at every stage. Autonomy is directly linked to the system’s ability to infer outputs from data.
3. Ability to adapt (optional criterion)
An AI system may adapt its behavior after deployment through machine learning or self-adjustment mechanisms. However, this ability is not mandatory for a system to be qualified as AI: a non-adaptive system may still fall within the definition as long as it satisfies the other criteria.
4. Explicit or implicit objectives
AI systems are designed to achieve specific purposes:
Explicit objectives are directly encoded by the designers (e.g. maximizing a cost function);
Implicit objectives may emerge from the system’s learning or interactions with the environment. These objectives should not be confused with the intended purpose, which refers to the actual use of the system in a given context.
5. Inference of outputs (learning or reasoning process)
A central criterion: the system uses AI techniques (supervised, unsupervised or reinforcement learning, symbolic logic, knowledge-based approaches, etc.) to infer patterns or conclusions from input data. This inference process takes place both:
during the design phase (training, modeling),
and during the operation phase (generation of outputs).
6. Production of specific outputs
AI systems generate outputs that fall into one or more of the following categories:
Predictions: estimating unknown values from known data;
Content: generating text, images, music (e.g. generative AI);
Recommendations: suggesting actions or options based on preferences;
Decisions: automating choices previously made by humans.
7. Influence on environments
AI systems produce concrete effects in the physical environment (e.g. mechanical action, robotics) or virtual environment (e.g. data flows, digital decisions). The system is therefore not passive, but acts on its environment through its outputs.
If the AI meets these criteria, then it becomes an AI system within the meaning of the AI Act and must comply with its requirements. Otherwise, the Regulation does not apply.
Components of an AI system:
An AI system is a broader and more complex application that integrates one or more AI models to perform a specific task. It includes not only the AI models, but also the components needed to collect, process and analyze data, as well as to interact with users.
In other words, an AI system is a complete solution that implements AI models within an operational framework.
- AI models: algorithms trained to perform predictions or analyses.
- Data collection and processing: processes used to gather and prepare data for the model.
- Infrastructure: the hardware and software needed to run the system, such as servers and databases.
- User interface: the means by which users interact with the system, such as web or mobile applications.
Example of an AI system:
- Virtual assistant: such as Siri or Alexa, which uses several AI models for speech recognition, natural language understanding, and response generation, while integrating databases and user interfaces to interact with users.
- Recommendation systems: used by platforms such as Netflix or Amazon to suggest content or products, integrating collaborative filtering models and user data processing.
Distinction between AI systems and traditional software
Recital 12 explains that the definition of AI systems should distinguish these systems from software systems or simpler traditional programming approaches.
The key criterion for distinguishing an artificial intelligence system (AIS) from traditional rule-based software is the ability to infer.
While conventional software automatically performs operations based on rules predetermined by humans, AI systems are designed to determine on their own (or “infer”) how to generate outputs from the input data received.
This distinction is based on several specific criteria:
• Source of the rules and applied logic
Traditional software relies on fixed rules coded exclusively by natural persons, in order to perform specified tasks. AI systems, by contrast, use techniques such as machine learning, logic-based approaches or symbolic knowledge-based approaches, enabling them to extract rules or patterns from data or from a coded knowledge base.
• Presence of learning and modeling mechanisms
Conventional data-processing systems do not include any learning, reasoning or modeling phase throughout their lifecycle. An AI system, by contrast, is characterized by the construction of internal models, often developed during a specific training or design phase, which enables it to solve complex tasks not defined by fixed instructions.
• Ability to handle complexity
Traditional software is designed for simple and stable tasks, such as sorting data, using spreadsheets or descriptive analysis. AI systems stand out for their ability to process complex relationships and dynamic data patterns, allowing them to evolve in uncertain and changing environments (e.g. autonomous driving, speech recognition).
• Degree of operational autonomy
An AI system operates with a variable level of autonomy, meaning with a certain degree of independence from human intervention. Traditional software follows a rigid “human input – determined output” sequence, whereas an AI system can produce results not explicitly predefined by a human.
• Nature of the heuristics applied
Systems based on classical heuristics (rules of thumb or standard algorithms such as MiniMax in chess) are excluded from the definition of an AI system. These approaches, while automated, display neither data-based learning capability nor adaptability.
• Level of sophistication of predictions
An automated system using basic statistical rules (e.g. calculating a historical average to predict prices or weather) does not constitute an AI system. AI systems go further by identifying complex and non-linear correlations, going beyond simple statistical models or conventional optimization rules.*
A system that performs operations exclusively according to “if… then…” rules written by a human, without any capacity for data modeling or autonomous inference, must be considered traditional software and not an artificial intelligence system within the meaning of the Regulation.
Map your AI systems with Dastra
Mapping enables companies and organizations to visualize all deployed AI systems, identify applicable legal obligations, and prioritize compliance actions.
For AI use-case mapping to be truly operational, it must go beyond a simple descriptive list and become a structured repository linked to the reality of systems, data and regulatory obligations. DASTRA addresses this need precisely by offering features that make it possible to document, classify and track AI systems in their organizational and regulatory context.