Software that does not meet the Article 3(1) definition is not an AI system for the purposes of the Act. That includes, in outline, purely deterministic software that applies fixed human-written rules without inference, simple calculation and lookup tools, and human processes that use software only to record or display information. The conclusion must be evidenced from how the software actually produces its outputs. Falling outside the AI Act does not mean falling outside every law, and a later change that adds inference can bring the software inside the definition.
What does falling outside the definition look like?
A system falls outside Article 3(1) when it lacks one of the defining elements. The clearest cases are tools that apply a fixed rule set written in full by a human, where the same input always produces the same output through a predetermined calculation, and no inference step derives how to generate the output. A tax calculator that applies published rates, a form validator that checks required fields, and a dashboard that displays data exactly as stored are illustrations of that pattern. The illustration only helps if the real software matches it. A calculator that also predicts, ranks or recommends using a learned model has added an inference layer, and the analysis must then separate the deterministic part from the inferential part.
How should a not an AI system conclusion be evidenced?
A defensible conclusion names the software version, describes the processing steps in order, and states which element of Article 3(1) is absent and why. Evidence can include design documentation, source-level descriptions of the rule logic, supplier statements that have been checked, and observed behaviour on test inputs. A single assertion that a tool is rule-based is not evidence. The record should also note who made the assessment, what was examined, and what change would trigger a reassessment, such as adding a machine learning model, connecting a general-purpose model, or introducing adaptive behaviour. Without that trigger, a later upgrade can silently move the software inside the Act.
For the broader regulatory framework, see EU AI Act fundamentals.
What about software that only assists a human decision?
Human involvement does not by itself take software outside the definition, and human involvement does not by itself bring software inside it. The test remains inference, autonomy and output influence. A spellchecker that suggests corrections the user may accept or reject still needs the inference test applied to how suggestions are generated. A workflow tool that routes a document to a human reviewer, without generating predictions, content, recommendations or decisions, is more likely to fall outside the definition, provided it does not infer anything beyond applying fixed routing rules. In each case, describe what the software generates and what the human does with it, rather than relying on the phrase human in the loop.
What about traditional analytics and statistics?
Traditional analytics that summarise past data using fixed formulas, without inferring how to generate new predictions, content, recommendations or decisions, may fall outside the definition. Analytics that predict a future outcome, score a person, classify an input, or generate a recommendation involve inference and need the full test. The dividing line is not the word analytics or the age of the method. Regression, clustering and similar statistical techniques can involve inference when they generate outputs that influence an environment. A record that only names the technique, without describing the output and its influence, does not answer the question.
Which other laws can still apply?
Software outside the AI Act can still be governed by data protection law where personal data are processed, by consumer law where it is supplied to consumers, by product safety and sectoral law where it forms part of a regulated product, by employment law where it affects workers, and by contract and liability rules. A conclusion that the AI Act does not apply should therefore be recorded as a conclusion about one regulation, not as a general statement that the software is unregulated. Cross-referencing the other assessments, or recording that they were considered separately, prevents a narrow AI Act conclusion from being read as a broader compliance claim.
Continue with what counts as an AI system.
What is left open by this guide?
This guide gives illustrations and an evidence method. It does not decide that any particular software product is outside the Act. Hybrid products, supplier updates, and changes in how outputs are used can all move a product across the definitional line, and the line itself will be shaped by guidance and practice over time. A team that is unsure, or that faces a borderline hybrid, should keep the question open in the record, obtain the technical evidence that would close it, and confirm the position with a competent authority or legal adviser where the deployment affects people in consequential ways. Keeping the question open is not indecision: it is the record showing that the definitional test was applied honestly, that the missing evidence was named, and that the system was not cleared on an assumption while the facts were still incomplete.
Frequently asked questions
Is a spreadsheet an AI system?
A spreadsheet that only applies fixed formulas entered by a user does not meet the Article 3(1) definition on its face. If a spreadsheet or an add-in uses a learned model to predict, classify or generate content, that feature needs its own assessment.
Does software need to learn after deployment to be an AI system?
No. Adaptiveness after deployment is described as a possibility in Article 3(1), not a requirement. A system can be an AI system even if its behaviour does not change after deployment, provided the other elements are present.
If software is not an AI system, is no record needed?
A short record is still valuable. It shows the version assessed, the reasoning, the evidence examined and the trigger for reassessment, so that a later reviewer can see why the conclusion was reached and when it should be revisited.
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 ↗