Every deployer of AI agents under the EU AI Act needs the same core items: an inventory with a risk classification per agent, staff AI literacy, disclosure to people who interact with agents, and a check that no agent can carry out a prohibited practice. High-risk agents add oversight, input control, monitoring, logs and notices.
What does the Act require of a deployer, in plain words?
The EU AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024 and applies in phases. Prohibitions applied from 2 February 2025 and general-purpose model obligations from 2 August 2025. Dates for high-risk obligations have been subject to proposed changes, so build to the substance rather than a date.
The Act is risk-based. Duties depend on whether a system is prohibited, high-risk, subject to transparency duties, or minimal risk, and on whether you are its provider or its deployer. A deployer's duties are lighter than a provider's, but they are not light, and they cannot be handed to the vendor. For background, see what the EU AI Act is.
Who should use this checklist?
Any business that uses AI agents in the EU, or whose agents' output is used in the EU, acting as deployer. If you build agents yourself or substantially change bought ones, you may also be a provider and carry the heavier set, which this checklist does not cover in full.
The checklist is organised by agent, not by firm. One agent may be high-risk in one capability and minimal risk in another, so classify the capabilities first. A deeper walk through that step is in our EU AI Act risk classification checklist.
Why does an agent change the checklist?
Most AI compliance lists were written for models that produce a score or a piece of text. Agents act. They call tools, write to records and sometimes move money. That adds three items most lists lack: classification by capability, oversight set at the moment of action, and a record that captures actions rather than conversations. The argument is developed in the EU AI Act and agents that act.
What is on the checklist, and what evidence does each item need?
The table runs from duties that apply to every agent to those that apply only to high-risk ones.
| Requirement | Applies to | Control | Evidence it leaves |
|---|---|---|---|
| Inventory and classification | All agents | Register of every agent and capability, classified by risk tier, with provider or deployer role stated | Signed register, reviewed on change |
| No prohibited practice | All agents | Capability review against the banned practices; blocks on manipulation and exploitation of vulnerability | Review record; tests showing refusals |
| AI literacy | All agents | Role-based training for briefers, approvers, overseers and reviewers | Training and assessment records by role |
| Tell people they are dealing with AI | Agents that interact with people | Disclosure at the start of each interaction, in every channel | Disclosure text and channel screenshots |
| Mark AI-generated content | Agents that generate certain content | Marking or labelling where the Act requires it | Marking method and samples |
| Use per instructions | High-risk | Configuration within the provider's stated purpose | Configuration mapped to the instructions version |
| Human oversight | High-risk | Named, trained overseers with authority to act; approval before adverse or irreversible actions | Assignments, approvals, override rates |
| Input data | High-risk | Source register with owner and trust class | Register; source per action in the record |
| Monitoring | High-risk | Review of refusals, holds, overrides and errors; tested stop control | Monitoring reports; suspension log |
| Logs | High-risk | Append-only action record kept for the required period | Retention schedule; integrity checks |
| Serious incidents | High-risk | Route to provider and authority | Incident tickets with notification times |
| Inform workers | High-risk used at work | Notice to workers and representatives before use | Notices and minutes |
| Inform affected people | High-risk deciding about people | Notice in decision communications | Templates and samples sent |
| Explanation on request | High-risk deciding about people | Process to explain the agent's role in a decision | Request log and responses |
| Rights impact assessment | Public bodies; credit scoring and life or health insurance pricing | Assessment before first use, notified to the authority | Assessment and notification receipt |
| Data protection impact assessment | Where GDPR requires one | DPIA using the provider's information | Signed DPIA |
Two items deserve emphasis. Oversight is the one supervisors test hardest, by asking overseers what the agent can do and looking at how often they disagree with it. Logs are the one most often inadequate, because they record what the agent said rather than what it did.
What will a supervisor or auditor ask?
- Show us the register. How did you classify this agent, and who signed it off?
- Which capabilities are high-risk? Show the oversight for those.
- Pick a case from last month and walk us through every action.
- How do people know they are dealing with an AI agent?
- How are staff trained, and how do you know it worked?
- Who in your organisation answers for the agents?
Enforcement sits mainly with national market surveillance authorities and, for financial firms, often with existing financial supervisors. Our explainer on who enforces the EU AI Act for AI agents covers the split.
What does the checklist not cover?
It does not cover provider duties such as conformity assessment, technical documentation and quality management. It does not cover GDPR beyond the DPIA link, and personal data processing by agents raises its own duties. It does not cover sector law: conduct rules, model risk, DORA for ICT risk in financial entities, or AML rules. And it does not cover your own risk appetite, which should set limits for agents that the Act never mentions. For the wider set of rules and the control that meets each, 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.
Where Pryme fits
Pryme keeps the evidence this checklist asks for as a by-product of running agents. Each agent carries a certificate from Attest and a depth; every action meets the minimum chain of five bounds; and every verdict, allowed, held, refused or withheld, is filed in an append-only record that names the bound that decided.
Questions people ask
Do minimal-risk agents need anything?
Yes. AI literacy applies to all AI systems, prohibited practices are banned for all, and any agent that interacts with people needs disclosure. Those three items apply even to a low-stakes internal agent.
How often should the checklist be rerun?
Whenever an agent gains or loses a capability, changes its sources, or its use changes, and on a fixed cycle otherwise. Classification can change with a single new connection.
Can one person own EU AI Act compliance for all agents?
One person can coordinate it. Each agent should still have a business owner who answers for its use, because the duties of oversight and monitoring sit where the agent works.