AI risk classes: tiers, duties and classification
Definition
AI risk classes are the categories into which the EU AI Act sorts AI systems by their potential for harm. From banned practices through high-risk systems and applications with transparency duties to minimal risk, the scope of duties rises with the class. The correct assignment determines which rules apply to a concrete system.
Why the Term Matters
Whether an AI project involves little or much effort depends almost entirely on the risk class. Whoever misjudges it either plans too tightly and stumbles over duties, or too cautiously and burns budget. As a neutral advisor we classify a system cleanly before it is built. A yes-or-no question becomes a solid basis for time, cost and feasibility.
When It Is Relevant
- An AI project is being planned and the scope of duties is still unclear.
- A system touches sensitive areas such as hiring, credit, access or safety.
- Legal risk should be estimated before an investment decision.
- An application might fall into the high-risk area and needs clarity.
When Not
When a system obviously supports only uncritical, internal tasks, a short assignment without deep analysis suffices. And when the use case is not yet decided, a risk-class framing is premature because it can still change with the concrete design.
Industries That Use It
- Mechanical engineering: AI in safety components can count as high-risk and needs early framing.
- Staffing providers: selection and scoring systems often fall into the high-risk area.
- Credit industry: creditworthiness scoring is a classic high-risk application with clear duties.
- Telecommunications: systems with chatbot or recognition functions trigger transparency duties.
Typical Use Cases
- Assigning each planned system to a risk class.
- The resulting list of duties per system.
- A check whether a project falls into a banned area.
- An estimate of effort and feasibility based on the class.
- Documentation of the assignment as a basis for governance and audit.
Decision Guide
| Signal | Risk-class framing fits | Alternative |
|---|---|---|
| Scope of duties unclear | classify before planning | planning blind is risky |
| Sensitive application area | precise assignment needed | underestimating gets costly |
| Clearly uncritical and internal | a short assignment suffices | deep analysis would be overbuilt |
| Use case still open | make it concrete first | the class may still change |
Example
A machine builder plans AI in a safety function. We check whether the system counts as high-risk, derive the resulting duties and estimate effort and deadlines. A second, internal tool falls into a low class with simple documentation. This describes the method, not an asserted client metric.
Service Connection
The entry is usually EU AI Act consulting, often within the AI potential analysis. For delivery, AI implementation follows.
Sources
This framing describes the logic of the risk classes in the EU AI Act. Statements on classes, duties and bans are general orientation and not legal advice; they must be checked case by case against the currently applicable legal texts and guidance.