Classify each use, not the agent as a whole. List every capability and its purpose, check each against the EU AI Act's prohibited practices, then its high-risk areas such as credit scoring of individuals, then its transparency duties for systems that interact with people. Record your role, provider or deployer, and the reasoning, and reclassify when a capability changes.
What are the EU AI Act's risk categories?
The EU AI Act, Regulation (EU) 2024/1689, takes a risk-based approach. It sorts AI uses into four broad groups: prohibited practices, which may not be placed on the market or used at all; high-risk systems, which carry the heaviest obligations; systems with transparency obligations, where people must be told they are dealing with AI or that content is AI-generated; and minimal-risk systems, with no specific obligations under the Act.
The Act entered into force on 1 August 2024, with obligations phasing in over several years. The prohibitions applied from 2 February 2025. Application dates for the high-risk obligations have been subject to proposed changes, so check the current position before you plan around them.
The overview of how these rules map to controls for agents is in our guide to AI agent regulation, mapped to controls.
Why classify per capability rather than per agent?
The Act attaches to the purpose for which an AI system is used. An agent that drafts customer emails, summarises complaints and assesses whether an individual can afford a loan is doing three things with three different purposes. The first two may carry only transparency obligations. The third, creditworthiness assessment of a natural person, is listed in Annex III as high-risk.
If you classify the agent as a whole, you either over-classify everything it does, which buries the team in obligations that do not apply, or under-classify the one capability that matters. The general method of classifying AI agents by risk applies here, with the Act's categories as the scale.
What are the steps?
- Agent owner: lists every capability and states its purpose in one sentence: what decision or output it supports, for whom, and with what effect.
- Compliance: checks each purpose against the prohibited practices. If any capability falls within them, it stops. There is no compliance route for a prohibited use.
- Compliance: checks each purpose against the high-risk areas. Annex III lists, among others, creditworthiness assessment and credit scoring of natural persons, and risk assessment and pricing for life and health insurance. Read the list in full for your sector; some uses by public bodies also appear.
- Compliance: checks transparency obligations. Article 50 requires that people are told when they interact with an AI system, and that certain AI-generated content is marked. Most customer-facing agents meet this test.
- Compliance: records anything left as minimal risk, with the reason.
- Legal: determines your role for each capability. Providers and deployers have different obligations. A firm that builds or substantially changes an agent for its own use may take on provider duties.
- Agent, depth 2: drafts the classification record from the capability list and the decisions above.
- Compliance lead: signs the record and sets the triggers for review.
- Agent owner: reclassifies whenever a capability is added, its purpose changes, or it starts affecting a new group of people.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| List capabilities | Lists its capabilities from its bounds | 1 | Stating purposes |
| Prohibited | Nothing | none | Compliance |
| High-risk | Nothing | none | Compliance |
| Transparency | Nothing | none | Compliance |
| Minimal | Nothing | none | Recording reasons |
| Role | Nothing | none | Legal |
| Record | Drafts | 2 | Checking |
| Sign | Nothing | none | Compliance lead |
| Reclassify | Flags new capabilities | 1 | Agent owner |
A question-by-question version is in the EU AI Act risk classification checklist, and the detailed question of whether an agent is high-risk is covered in is an AI agent a high-risk AI system under the EU AI Act?.
Which controls make the classification safe?
Purpose written per capability. Evidence: the capability list with purposes.
Classification signed by a named person. Evidence: the signed record.
Reasoning recorded, not just the result. Evidence: why each capability is or is not in each category.
Triggers for review. A new capability, a widened depth that changes the effect on people, a new customer group. Evidence: the trigger list and each review.
Link to risk management. Where a capability is high-risk, it enters the risk management process the Act sets out in Article 9; see EU AI Act Article 9 risk management for AI agents. Evidence: the risk register entry.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What are the signs classification is going wrong?
The vendor's classification is adopted. A supplier says its product is "not high-risk". That may be true for the product in general and false for how you use it. The classification follows your purpose.
A person in the loop is treated as an exemption. Having a person approve the agent's output may be part of meeting high-risk obligations; it does not by itself take a listed use out of the high-risk category.
Capabilities are added without review. An agent that started by drafting emails is given a credit assessment capability, and the classification record still says "transparency only".
Depth is ignored. A capability that only reads and summarises is used differently from one that acts. Widening depth can change what the capability does to people, and should trigger a review.
Nothing is written down. The team believes the agent is minimal risk but cannot show why. An auditor will ask for the reasoning.
What is outside the classification?
The Act's risk categories are not the only rules that apply. Data protection law, including GDPR Article 22 on decisions based solely on automated processing, sector conduct rules, and operational resilience rules such as DORA apply separately. A capability classified as minimal risk under the AI Act may still face strict obligations elsewhere.
Where Pryme fits
In Pryme, capabilities are defined separately, each with its own purpose and depth, which makes per-capability classification practical. Adding a capability is a change that appears in the agent's record, so it can trigger a classification review.
Questions people ask
Can one agent have capabilities in different categories?
Yes, and most do. Classify each, and apply the obligations for the highest category to that capability, not to the whole agent.
Does an internal-only agent need classifying?
Yes. Classification follows purpose, not audience, and an internal capability can still fall in a listed area. Classify every capability, internal or external.
How often should the classification be reviewed?
At every trigger, and at least once a year, since the agent, its uses and the guidance around the Act all change.