Do not be misled by the labels. Terms like high risk, limited risk and minimal risk sound narrow and technical, but in practice they cover a wide range of real systems. Many AI services operating today fall into these categories immediately, including systems that appear simple, harmless or purely assistive.
Most developers are swamped by superficial compliance and governance noise. None of it helps you understand where your system actually sits. Categories matter because they determine what AI behavioural testing evidence you must produce.
This page explains what the categories mean in operational terms so you can identify where your system sits and what obligations follow.
It does not matter whether you operate a local AI system or rely on a service provided by a large company. The responsibility sits with you, and the obligations follow your use of the system, not the size or reputation of the provider behind it. Your category determines what AI behavioural testing evidence you must produce to show you understand and control how your system behaves.
Your category determines:
Misclassification is one of the fastest ways to fall out of compliance. Most developers rely on templates, frameworks and governance material instead of testing how their system behaves, which makes misclassification even more likely.
These systems are banned outright because they present unacceptable risks. Examples include:
If your system falls here, it cannot be placed on the EU market under any circumstances. No amount of documentation, governance material or testing can change this classification.
These systems are allowed but heavily regulated. High risk systems include:
High risk systems must meet strict requirements covering data quality, documentation, AI behavioural testing evidence, monitoring and human oversight. This category captures far more systems than most teams expect, especially when they rely on superficial compliance and governance material instead of testing how their system behaves under real conditions.
These systems must meet transparency requirements. Examples include:
Users must be informed they are interacting with AI and given clear instructions for safe use. Transparency is not administrative paperwork. It is an operational requirement that depends on understanding how your system behaves and where it may mislead. AI behavioural testing evidence is the only reliable way to demonstrate this.
Most everyday AI falls here. Examples include:
Minimal risk systems have no mandatory obligations, but you still need to understand how they behave. Good practice means having AI behavioural testing evidence so you can show the system works as intended and does not mislead users.
These are large scale AI systems trained on broad data and used for many downstream tasks. They must meet:
This applies whether you are the original developer or a downstream deployer adapting the system. Obligations follow your use, and you must be able to show behavioural evidence for how the system performs in your context.
General purpose AI covers systems that can be used for many tasks across different contexts. They require:
If a general purpose system is integrated into a high risk context, the high risk rules apply. Obligations follow your use, and you must be able to show behavioural evidence for how the system performs in that context.
This category covers the largest and most capable AI systems with the potential to cause widespread harm if misused. They require:
This category is reserved for systems with exceptional scale and impact. You must be able to show behavioural evidence for how the system performs, fails and recovers under real conditions.
You can classify your system by checking:
Ignore governance noise. Classification depends on how your system behaves, not what your documentation claims. If you are unsure, start with the Check if the EU AI Act Applies to You page and follow the workflow.
Developers are surrounded by compliance noise that does not matter when regulators ask what the system actually does. You need AI behavioural testing evidence that shows how your system behaves under real conditions.
These resources give you a clear starting point for understanding your obligations and producing behavioural evidence that stands up when questioned.
Test to Comply Now(C) William Argo