AI risk classes: definition, tiers and classification
Definition
AI risk classes are the categories the EU AI Act uses to sort AI systems by their potential for harm. Broadly they range from banned practices through high-risk systems and systems with transparency duties to minimal risk. The class determines which duties apply. For an AI assistant the classification is the first step, because it sets the entire effort.
Why the Term Matters
Without classification nobody knows whether an assistant needs just a label or full documentation and oversight. The risk class translates an abstract legal text into a concrete to-do list. Who determines it early avoids two costly mistakes: over-regulating a harmless assistant and underestimating a critical agent.
When It Is Relevant
- A new AI system is to go into productive use.
- An assistant touches sensitive areas like selection or credit.
- You want to plan the compliance effort realistically.
- A bought-in tool must fit your own classification.
When Not
A pure lab experiment with no bearing on people does not yet need a formal classification, though it does need a rough first estimate. And the class alone solves nothing: it says which duties apply but does not replace implementing them with rights, approvals and logs.
Industries That Use It
- Recruitment: selection systems often fall into high-risk and need strict obligations.
- Lending institutions: credit scoring regularly sits under special scrutiny.
- Critical infrastructure: control systems are classified strictly.
- Retail and marketing: a pure info assistant usually stays in a low class.
Typical Use Cases
- Classify a company brain as a minimal-risk system.
- Put a chat assistant under the transparency duty.
- Check an HR agent against high-risk criteria.
- Rule out banned practices early, before building.
- Sort a bought-in solution into your own class model.
Decision Guide
| Signal | High class likely | Low class likely |
|---|---|---|
| Decision about people | check high-risk | pure info is low |
| Sensitive area affected | strict obligations | neutral factual topics lighter |
| Direct chat with users | at least transparency duty | internal use often lower |
| Manipulative patterns planned | ban looms | factual help is uncritical |
Example
Before building, a team classifies: the planned company brain only returns internal knowledge and lands in minimal risk, the outward chat assistant gets an AI label, and a planned selection agent is deferred as a possible high-risk case and reviewed. This describes the method, not legal advice and not an invented metric.
Service Connection
We classify every system early and build a company brain, an AI assistant or AI agents to fit the respective risk class, with labelling, rights and approvals.
Sources
This framing describes the rough structure and is not legal advice. The exact criteria and boundaries of the risk classes follow from the EU AI Act and should be checked against the current regulation text and official guidance case by case.
