Prohibited, high-risk, transparency and minimal risk: those are the EU AI Act's four tiers, and they are set by use, not by technology. Most business AI agents land in transparency or minimal risk. Agents that assess individuals' creditworthiness or price life and health insurance land in high-risk. None may ever carry out a prohibited practice.
What are the four tiers?
The EU AI Act (Regulation (EU) 2024/1689) is risk-based. It sorts AI systems by what they are used for:
| Tier | What it covers | What it means for the system |
|---|---|---|
| Prohibited | Practices banned outright, such as manipulation causing significant harm, exploiting vulnerability, social scoring, and certain biometric uses | May not be placed on the market or used. Applied from 2 February 2025 |
| High-risk | Systems used as safety components of regulated products, and uses listed in Annex III, including creditworthiness assessment and credit scoring of individuals, and risk assessment and pricing in life and health insurance | Full provider duties (risk management, data governance, logging, transparency, oversight, accuracy) and deployer duties |
| Transparency | Systems that interact with people, or generate certain content | People must be told they are dealing with AI; certain content must be marked |
| Minimal risk | Everything else | No specific duties beyond AI literacy, which applies to all |
General-purpose AI models sit alongside this structure with their own obligations, which have applied since 2 August 2025. Dates for high-risk obligations have been subject to proposed changes, so build to substance rather than a date.
Who decides which tier an agent is in?
The provider classifies the system for its intended purpose. The deployer must use it within that purpose, and if the deployer changes the purpose so that the system becomes high-risk, the deployer may become its provider. In practice, every deployer should classify its own agents, because the deployer controls how they are actually used.
The Act also lets a provider show that a system listed in Annex III is not high-risk where it does not pose a significant risk of harm, for example because it performs a narrow procedural task or only prepares an assessment for a person. That reasoning must be documented, and it does not apply to systems that profile individuals. Our EU AI Act risk classification checklist walks through the tests.
Why do agents complicate the tiers?
The tiers assume a system has one purpose. An agent often has several capabilities, and each can sit in a different tier.
Take a lender's servicing agent. It answers balance questions (transparency tier, because it talks to customers). It drafts internal notes (minimal risk). It gathers documents for a top-up loan and summarises affordability for an underwriter (part of a high-risk process, because it contributes to assessing an individual's creditworthiness). One agent, three tiers.
Two consequences follow. First, classify by capability, then take the highest tier for the agent as a whole unless you can separate the capabilities cleanly. Second, watch for capability creep: one new connection can move an agent up a tier without anyone deciding it should. The step-by-step method is in how to classify an AI agent under the EU AI Act risk categories.
Where do typical business agents land?
- Customer support agent answering questions and handling routine requests: transparency. See whether a customer support chatbot is high-risk; usually not.
- Accounts payable agent matching invoices and drafting payments: minimal risk under the Act, though payment and fraud rules apply.
- Credit underwriting assistant reading statements and recommending decisions on individuals: high-risk.
- Life or health insurance pricing agent for individuals: high-risk.
- Fraud detection agent: the credit listing carves out detecting financial fraud, so usually not high-risk on that ground.
- Recruitment screening agent: high-risk, because employment uses such as screening and ranking candidates are listed.
- Collections agent: not listed, but at real risk of the prohibitions on manipulation and exploiting financial vulnerability.
- Internal research or drafting agent: minimal risk.
The question whether an AI agent is a high-risk AI system therefore has no single answer. It depends on what this agent does.
How do you meet the tiers in practice?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Classify each system | Register entry per capability with tier and reasoning | Signed classification memo |
| Avoid prohibited practices | Brief and platform restrictions on manipulation and exploitation | Restrictions and test results |
| Meet high-risk duties | Oversight, logs, monitoring, notices for high-risk capabilities | Duty-by-duty evidence file |
| Meet transparency duties | Disclosure in every channel | Disclosure captures |
| Stop tier creep | New connections and capabilities reviewed for tier before enabling | Change review records |
| AI literacy for all | Role-based training | Training records |
High-risk classification brings the Article 9 risk management system with it; see EU AI Act Article 9 risk management for AI agents.
What will a supervisor or auditor ask?
- Show us each agent's tier and the reasoning.
- If you decided a listed use was not high-risk, why?
- What has changed since you classified it?
- Which capabilities of this agent contribute to decisions about individuals?
What do the tiers not tell you?
They do not tell you how dangerous an action is. A minimal-risk payment agent can still send money to the wrong account; the Act will not help, but payment law, fraud rules and your own limits will. Nor do the tiers capture reversibility, which for agents is often a better guide to risk than the use category. Classify under the Act, then set controls by what each capability can do. For the full map, see AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Can an agent be in more than one tier?
Its capabilities can. The Act classifies systems by intended purpose, so a multi-purpose agent should be treated at its highest tier unless the capabilities are separated and governed apart.
Does using a powerful model put an agent in a higher tier?
No. The tier depends on use. Model capability matters for the general-purpose model rules, which apply to the model provider, not to your classification.
Is minimal risk the same as no duties?
No. AI literacy applies to every AI system, and prohibited practices are banned for all. Sector rules and data protection law apply whatever the tier.