The core logic of the EU AI Act is simple: the higher the risk, the stricter the rules. That's why the first step towards compliance is to determine which risk category each of your AI systems falls into. Classification happens not at organisation level but at system and intended-purpose level; the same model can be minimal-risk in one use and high-risk in another.
The four risk categories
1. Prohibited practices
Practices posing an unacceptable risk to fundamental rights have been banned since February 2025. Examples:
- Influencing people through subliminal or manipulative techniques in a harmful way
- Exploiting vulnerabilities such as age, disability or socio-economic situation
- Social scoring
- Emotion recognition in the workplace and in education (except for medical or safety reasons)
- Building facial recognition databases through untargeted scraping from the internet or CCTV
2. High-risk systems
Systems in areas that can significantly affect people's lives and rights. There are two groups:
- Annex III: recruitment and workforce management, creditworthiness assessment, life and health insurance pricing, education, access to essential services, biometrics, critical infrastructure, law enforcement, migration and justice. Obligations apply from 2 December 2027.
- Annex I: safety components of products covered by EU product safety legislation (medical devices, machinery, toys, vehicles, etc.). Obligations apply from 2 August 2028.
3. Systems with transparency obligations
Systems that interact with people or generate content. Users must be told they are interacting with AI, generated content must be marked, and deepfakes must be disclosed. These obligations apply from August 2026.
4. Minimal-risk systems
Systems such as spam filters, inventory forecasting and most productivity tools. No additional obligations; voluntary codes of conduct are encouraged.
One obligation applies across all categories: organisations in scope must support their staff's AI literacy (Article 4).
Classification examples by sector
| System | Likely category | Note |
|---|---|---|
| Consumer credit application scoring model | High risk | Annex III, creditworthiness of natural persons |
| Card fraud detection model | Not high risk | Financial fraud detection is explicitly excluded |
| Bank customer service chat assistant | Transparency | Users must be told it's AI |
| System that screens CVs and ranks candidates | High risk | Annex III, recruitment |
| System that evaluates employee performance | High risk | Annex III, workforce management |
| Call-centre system analysing employees' emotions | Prohibited | Emotion recognition at work |
| Tool generating marketing images and copy | Transparency | Generated content marked; deepfakes disclosed |
| Life insurance premium pricing model | High risk | Annex III |
| Internal document search and summarisation assistant | Usually minimal or transparency | Depends on intended purpose |
| Demand and inventory forecasting model | Minimal | Makes no decisions about people |
The high-risk exception
Even when a system operates in an Annex III area, the Act may not treat it as high-risk if it doesn't pose a significant risk to health, safety or fundamental rights and, for example, it:
- only performs a narrow procedural task,
- improves the result of a previously completed human activity,
- detects decision-making patterns without replacing human assessment,
- or only performs a preparatory task.
However, a system that profiles natural persons is always considered high-risk. Providers relying on the exception must document their assessment.
Your role matters too
For a system in the same category, obligations differ by role:
- Provider (builds the system or offers it under its own name): risk management, data governance, technical documentation, conformity assessment, registration and post-market monitoring
- Deployer (uses the system): use according to instructions, human oversight, keeping logs, informing people and, in some areas such as credit scoring, a fundamental rights impact assessment
If you substantially modify a system you've bought, or offer it under your own name, you can become a provider rather than a deployer.
Find out your system's category
MLTeam's free EU AI Act Scope Checker shows in a few questions whether your system is in scope, its likely risk category, your role and your obligations under the current timeline. We recommend running it separately for each AI system.
This article is for general information and is not legal advice.
Sources