Work through this EU AI Act risk classification checklist once for every AI agent, in order: is it in scope, is the use prohibited, does it fall in a high-risk area, does it carry transparency obligations, or is it minimal risk? Record the answer to each question with the reasoning and the evidence. The classification decides which obligations apply.
Why classify agents one by one?
The EU AI Act is risk-based. Regulation (EU) 2024/1689 sorts AI systems into four tiers: prohibited practices, high-risk systems, systems with transparency obligations, and minimal risk. The tier depends on the use, not the technology. Two agents on the same platform can sit in different tiers because one summarises internal documents and the other supports creditworthiness assessments of individuals. The tiers are explained in EU AI Act risk tiers explained for AI agents. How each tier maps to controls is covered in AI agent regulation, mapped to controls.
Agents also change use more often than traditional systems. A capability added to an existing agent can move it into a higher tier. Classification has to follow each change.
When should you use it?
At design, before any build, so the obligations shape the agent. Again before go-live, and at every change in use or capability. The step-by-step method is in how to classify an AI agent under the EU AI Act risk categories.
The result feeds two further checklists. If the agent is high-risk and you are its deployer, work through the deployer obligations checklist. For the full picture across tiers and roles, use the EU AI Act compliance checklist for AI agents.
How do you answer each question?
- Scope asks whether the agent is an AI system used in the EU, or whose output is used in the EU. Most agents built on modern models will be.
- Prohibited asks whether the use matches any practice the Act bans outright. Check the use against the Act's list with counsel. Regulated firms rarely meet this tier, but record that you checked.
- High-risk asks whether the use falls within an area listed in Annex III, or whether the agent is a safety component of a regulated product. For financial services, two Annex III areas matter most: creditworthiness assessment and credit scoring of natural persons, and risk assessment and pricing for life and health insurance.
- Transparency asks whether people interact with the agent directly, or whether it generates content that must be marked. Article 50 then applies.
- Supporting role asks whether the agent decides or only prepares. Record it, because it affects how you describe the use and, in some cases, how the classification is argued. Do not assume a supporting role takes a use out of a high-risk area; that is a question for counsel. Whether a credit scoring agent is high-risk is examined in is a credit scoring agent high-risk under the EU AI Act?.
The checklist
| No. | Question | Answer | Reasoning | Evidence |
|---|---|---|---|---|
| 1 | Is the agent an AI system within the Act's scope? | |||
| 2 | Is it used in the EU, or is its output used there? | |||
| 3 | Does the use match a prohibited practice? | |||
| 4 | Is the agent a safety component of a regulated product? | |||
| 5 | Does the use fall within an Annex III area? Which? | |||
| 6 | Does the agent decide, or prepare for a person to decide? | |||
| 7 | Do people interact with it directly? | |||
| 8 | Does it generate content that must be marked? | |||
| 9 | Resulting tier | |||
| 10 | Our role: provider, deployer, both | |||
| 11 | Reviewed by counsel? | |||
| 12 | Next review trigger |
A filled sample
A hypothetical bank, two agents.
| No. | Agent A: Personal loan pre-assessment | Agent B: Internal policy assistant |
|---|---|---|
| 1 | Yes | Yes |
| 2 | Yes, used in two member states | Yes |
| 3 | No; checked against the Act's list | No |
| 4 | No | No |
| 5 | Yes: creditworthiness assessment of natural persons | No |
| 6 | Prepares a summary and indicative score; an underwriter decides | Answers staff questions about internal policy |
| 7 | No; used by staff | Staff only |
| 8 | No | No |
| 9 | High-risk | Minimal risk |
| 10 | Deployer | Deployer |
| 11 | Yes, for Agent A | Not required |
| 12 | Any change to scoring inputs or a move to customer-facing use | Any use outside staff policy questions |
Row 12 matters for Agent B. If the bank later lets customers use it, row 7 changes and Article 50 applies. Writing the trigger down makes that change hard to miss.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Can a single agent fall into more than one tier?
The obligations can stack. A high-risk agent that people interact with directly carries both high-risk and transparency obligations. Record every tier that applies.
Does using a general-purpose model make our agent high-risk?
No. General-purpose AI models have their own obligations, which mostly fall on their providers. Your agent's tier depends on how you use it.
How often should classification be reviewed?
At every change in use or capability, and at least annually. Also review it when guidance from the Commission or national authorities changes how a use is read.