Why classification is the first step
The AI Act is built on a risk-based model: the higher the risk an AI system poses to people’s health, safety and fundamental rights, the stricter the requirements placed on you as a provider and deployer. Classifying AI systems correctly is therefore about first determining which role and which risk level each system has – that is the process we call classification.
Classifying your AI systems correctly is important both for knowing which measures are required to meet the legislation’s requirements and for avoiding unnecessary costs. A system that is incorrectly assessed as low-risk risks administrative fines if the supervisory authority later concludes that it should have been classified as high-risk. A system that is incorrectly classified as high-risk is instead burdened with unnecessary requirements for risk management, technical documentation and assessment of conformity with the legislation.
This article goes through the AI Act’s four risk levels, what determines whether a system is high-risk, and how you get started with classifying the AI systems in your organisation.
The four risk levels
The AI Act divides AI systems into four levels:
| Risk level | What it means | Examples | Requirements |
|---|---|---|---|
| Unacceptable risk | Prohibited use | Social scoring, manipulative AI, real-time biometric mass surveillance | Prohibited since 2 February 2025. |
| High risk | Full compliance required before the system may be put into service | Recruitment, credit assessment, biometric identification, education, critical infrastructure | Risk management, technical documentation, human oversight and conformity assessment apply from 2 December 2027. |
| Limited risk | Transparency towards the user | Chatbots, AI-generated content (deepfakes) | Information that one is interacting with or exposed to AI applies from 2 August and 2 December 2026. |
| Minimal risk | No specific regulation | Spam filters, AI in video games | Voluntary codes of conduct are recommended. |
Many AI systems on the market fall into the minimal or limited risk categories, and here there are certain requirements that need to be met, or other voluntary codes of conduct that are only a recommendation. It is the high-risk category, or close to high-risk, that requires the most work and where the line is hardest to draw.
Unacceptable risk – prohibited AI use
Since 2 February 2025, six types of AI use have been prohibited within the EU:
- Manipulative or deceptive AI that purposefully influences people’s behaviour or emotions in order to cause harm
- Exploitation of vulnerable groups linked to age, disability or socio-economic situation
- Social scoring – ranking people based on behaviour or socio-economic status
- AI that predicts the risk of committing crimes through predictive policing profiling based solely on personal traits or data patterns
- Emotion recognition in the workplace or at school
- Real-time biometric identification in public spaces, with exceptions for law enforcement under strict conditions
Thanks to the Digital Omnibus, which was voted through in July 2026, two further categories of prohibited AI use will be added on 2 December 2026:
- Nudify apps: AI apps that create unwanted and non-consensual undressed images (intimate deepfake material) are prohibited.
- Child sexual abuse: Tightened and expanded prohibitions on AI-generated material relating to child sexual abuse (CSAM)
If your organisation procures or builds AI solutions: screen suppliers against the list before contracts are signed. A system built for another purpose can still end up here if it can actually be used in a way that falls within any of the areas listed above. It is both the use and the original intent that are decisive.
In this article you can read more about prohibited AI use.
High risk
An AI system is classified as high-risk in one of two ways:
| Route | Legal basis | What is meant | New deadline |
|---|---|---|---|
| Product | Article 6(1) + Annex I | AI that is part of, or is a safety component of, a product already covered by EU harmonisation legislation (e.g. medical devices, vehicles, toys) and that requires third-party assessment. | 2 August 2028 |
| Use case | Article 6(2) + Annex III | Standalone AI systems used in specifically designated high-risk areas: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, administration of justice | 2 December 2027 |
Both deadlines were postponed through the EU’s “Digital Omnibus” by 12 and 16 months respectively compared with the original plan. The requirements are largely unchanged; it is only the point in time by which they must be met that has been moved back.
What makes a system a safety component?
In May 2026, the European Commission published draft guidelines on the classification of high-risk systems as an aid to making the assessment more predictable and consistent across all 27 member states. A safety component is a system that has any of the following functionalities:
- Malfunction: an error or failure in the system could endanger health, safety or property, even if that was not the intended purpose (e.g. AI that controls the timing of a lift door)
- Safety function: the system’s stated purpose is to prevent or mitigate risks to health, safety or property (e.g. collision detection in vehicles)
So, AI systems with functions for performance optimisation, comfort and efficiency do not count as safety functions and are therefore not to be regarded as high-risk systems under the AI Act, unless a failure in them creates a safety risk.
Intended purpose is decisive – document it carefully
The classification is governed by what the provider states the system is intended for or can be used for. What counts: instructions for use, technical documentation and marketing material. A general-purpose AI system without clear limitations risks being assessed on the basis of the use cases it could be used for, i.e. including high-risk ones. Third parties who change the intended purpose inherit the provider’s obligations.
Can you argue that your system, despite being used in a high-risk area, does not pose a risk to people’s health, safety and fundamental rights? Then you can claim that you may be exempted from the requirements for high-risk systems.
Limited risk – transparency requirements
Systems that interact directly with people or generate content are subject to transparency requirements under Article 50 of the AI Act, which apply from 2 August 2026. The transparency requirement covers the following areas:
- Chatbots and AI assistants must disclose that the user is interacting with AI
- AI-generated or AI-manipulated content such as images, audio and video (“deepfakes”) must be labelled as such
- Emotion recognition or biometric categorisation systems must inform the persons exposed to them
- AI-generated texts published to inform the public on matters of public interest must be labelled – unless the content has undergone human editorial review
The transparency requirements apply regardless of how the system is otherwise classified; it is the interaction with people itself that triggers the requirement, not the system’s overall risk level.
How to classify your AI systems – step by step
- Take inventory. List every AI system used in the organisation – purchased, built in-house or embedded in other software – on-prem, cloud, shadow AI.
- Determine your role for each system. Are you a provider who develops the system or places it on the market under your own name, or are you a deployer who uses a system developed by someone else? The obligations differ.
- Screen against the list of prohibitions. First rule out that the system falls within any of the prohibited practices.
- Test against Article 6 if you have a high-risk system. Is the system part of a product, or is it used within one of the high-risk areas in Annex III?
- Review the intended purpose. Go through the instructions for use and technical documentation. Does your area of use match what the system is intended for?
- Document the assessment. Even systems that end up in minimal or limited risk should have brief documentation saved. Supervisory authorities may ask for it.
- Set dates. Note which deadline applies to your particular system and link it to your internal compliance plan.
Summary and next steps
Most organisations already use more AI systems than they are aware of, and it is the classification that determines which compliance requirements await each system. So start the inventory and classification in good time. It lays the foundation for the rest of the work and makes it possible to prioritise the right systems first! This is extensive work with a lot of details to keep track of. In our platform AI Governance Starter, you get support for inventory, classification, risk assessment and action planning. Feel free to book a demo and we’ll tell you more!