AI system

On this page 7

An AI system is a machine-based system that infers from its inputs how to generate outputs such as predictions, content, recommendations or decisions that can influence a physical or virtual environment. In artificial intelligence, the term describes the whole operational unit that delivers a capability, not just the trained model inside it, and it is a central concept in AI fundamentals because regulation, engineering and risk management all attach to the system rather than to its parts.

The distinction matters in practice. A model is a component, while the system adds data pipelines, interfaces, rules, monitoring and human oversight. Two products built on the same model can therefore behave, perform and fail in very different ways, and responsibility for those differences sits at the system level.

How an AI system turns inputs into outputs

An AI system works by receiving input, applying an inference mechanism and acting on the result within a defined objective. The inference mechanism may be a learned model, a set of hand-written rules, a search procedure or a combination of these, which is why the definition focuses on inference rather than on a specific technique.

Most deployed systems share a recognizable set of components. The table below lists them with the role each one plays and a typical example from a production setting.

ComponentRole in the systemTypical example
Input layerCaptures and validates data from users, sensors or other softwareCamera feed, form fields, API request
PreprocessingCleans, normalizes and encodes inputs into the expected formatTokenization, image resizing, feature scaling
Inference componentProduces a prediction, score, plan or generated contentNeural network, decision tree, rule engine
Decision logicConverts raw outputs into actions using thresholds and business rulesApprove if fraud score is below 0.2
Output interfaceDelivers results to people or downstream systemsDashboard alert, chatbot reply, actuator command
Monitoring and feedbackTracks quality, drift and incidents over timeAccuracy reports, drift alarms, audit logs
Core components of a typical AI system

The objective may be explicit, such as minimizing a loss function during training, or implicit, as when a system learns preferences from examples. Either way, the objective is set by the people who design and operate the system, and an AI system does not choose its own goals in any meaningful sense.

Autonomy and adaptiveness as defining properties

AI systems differ chiefly in how much they act without human intervention and whether they keep changing after deployment. These two properties appear in the most widely used international definitions and strongly influence how a system must be governed.

The following table sets out common levels of autonomy, from advisory tools to systems that act directly on the world.

Autonomy levelHuman roleExampleMain governance concern
Decision supportHuman makes every decisionRadiology triage suggestionsAutomation bias in reviewers
Human in the loopHuman approves each proposed actionLoan pre-approval queueRubber-stamping under time pressure
Human on the loopHuman supervises and can interveneWarehouse robot fleetSpeed of intervention
Fully autonomousNo routine human involvementSpam filtering at scaleSilent failure and error accumulation
Levels of autonomy in AI systems

Adaptiveness refers to learning that continues in operation, for example a recommender that retrains nightly on new clicks. Adaptive systems can improve with use, but they also introduce feedback loops and drift, so their tested behavior at launch is no guarantee of their behavior months later.

An AI system is best understood by contrast with the AI model, the foundation model and the compound AI system, which are related but not interchangeable. Confusing them leads to errors in procurement, testing and legal compliance.

ConceptWhat it isRelation to an AI system
AI modelA trained mathematical function with fixed weightsOne component inside the system
Foundation modelA large general model adapted to many tasksA reusable base that many systems build on
Compound AI systemSeveral models and tools orchestrated togetherA specific architecture of an AI system
Expert systemKnowledge base plus inference engine using rulesA classic, rule-based kind of AI system
Conventional softwareExplicitly programmed logic with no inferenceNot an AI system under most definitions
How an AI system differs from neighboring concepts

The boundary with conventional software is the most contested. Simple spreadsheet formulas or fixed if-then scripts are usually excluded, whereas systems that derive their behavior from data or from logical inference are included. Regulators have acknowledged that borderline cases exist and are assessed case by case.

Why the system boundary matters in regulation

Legal frameworks regulate the AI system as the unit of accountability, because harm arises from how outputs are used, not from model weights in isolation. The OECD definition, revised in 2023, and the EU AI Act both describe a machine-based system that infers how to generate outputs, with varying levels of autonomy and adaptiveness.

Under the EU AI Act, obligations depend on the intended purpose and risk category of the system. The same image classifier could be minimal risk in a photo organizer and high risk in a hiring tool, because the deployment context changes the consequences of an error.

Risk categoryTypical useMain consequence
Unacceptable riskSocial scoring by public authoritiesProhibited
High riskRecruitment, credit scoring, medical devicesConformity assessment, documentation, oversight
Limited riskChatbots, generated mediaTransparency duties toward users
Minimal riskSpam filters, game opponentsNo specific obligations
Risk-based treatment of AI systems under the EU AI Act

Standards such as ISO/IEC 22989 for AI concepts and terminology and ISO/IEC 42001 for AI management systems also treat the AI system as the object of lifecycle management, from design and verification through operation and retirement.

Common failure modes and misconceptions

Many failures of AI systems originate outside the model, in data handling, integration or human use. Experienced practitioners therefore evaluate the full pipeline rather than a benchmark score alone.

  • Equating model accuracy with system quality: a model with high test accuracy can still fail when preprocessing in production differs from training.
  • Distribution shift: inputs change over time, and performance degrades without any visible error message.
  • Automation bias: operators accept outputs uncritically, so nominal human oversight provides little real protection.
  • Unclear ownership: when a system combines third-party models and in-house logic, responsibility for errors must be assigned explicitly.
  • Anthropomorphism: describing a system as understanding or wanting something obscures that it optimizes a defined objective.

Mitigation relies on end-to-end testing, continuous monitoring and documented fallback procedures. A well-engineered AI system degrades gracefully, for example by routing uncertain cases to a person instead of guessing.

In everyday language

An AI system is a complete working tool that looks at information and decides what to say or do next. The trained model is like the engine of a car. The AI system is the whole car: the engine, the steering, the brakes, the dashboard and the driver who watches the road.

This is why the same engine can end up in very different vehicles. What counts is the whole tool and how people use it, not only the clever part in the middle. When something goes wrong, the question is usually about the system as a whole.

Seen in the real world

A mid-sized hospital introduces an AI system to flag possible sepsis in ward patients. The system reads vital signs and laboratory values from the electronic health record every 15 minutes, a model computes a risk score, and decision logic sends an alert to the nurse in charge when the score passes a set threshold.

During the first month, nurses receive about 40 alerts per day, and many are false alarms caused by a laboratory feed that reports values in different units after midnight. The model itself is unchanged; the team fixes the unit conversion in preprocessing, raises the alert threshold slightly and adds a weekly review of missed cases.

After these changes, alerts fall to about 12 per day and staff begin to trust them. The outcome shows that reliability came from engineering the whole system, not from retraining the model.