How to meet deployer obligations for an AI agent

First confirm you are the deployer, not also the provider. Then, for each high-risk capability, use the agent within its instructions, assign oversight to named people with the competence and authority to intervene, control the input data you supply, monitor its operation, keep its logs, tell affected people where required and report serious incidents. Give each duty an owner.

Who is a deployer, and why does it matter?

Under the EU AI Act, a deployer is the organisation using an AI system under its own authority in a professional context. A provider is the organisation that develops the system, or has it developed, and places it on the market or puts it into service under its own name. The two carry different obligations. The definitions are explained in what a deployer is under the EU AI Act.

For agents, the line can blur. A firm that takes a platform and builds its own agent on it, writes the brief, chooses the sources and sets the purpose, may be a deployer of the platform and, depending on what it does, may take on provider duties for the agent it built. Settle this first, with counsel; the analysis is in who is the provider and who is the deployer of an AI agent.

The heaviest deployer obligations, set out mainly in Article 26, apply to high-risk systems. Other duties, such as AI literacy under Article 4 and transparency under Article 50, apply more widely. The full map is in AI agent regulation, mapped to controls.

What are the obligations, in substance?

For high-risk uses, Article 26 asks deployers, in substance, to:

• use the system in line with the provider's instructions for use;
• assign human oversight to people who have the competence, training and authority to carry it out, and support them;
• ensure that input data under their control is relevant and sufficiently representative for the intended purpose;
• monitor the system's operation, inform the provider of risks, and suspend use where a risk emerges;
• report serious incidents through the required channels;
• keep the logs the system generates, to the extent they are under the deployer's control;
• inform people where required, including workers and individuals subject to decisions made with the system's help.

Some deployers must also carry out a fundamental rights impact assessment under Article 27 before first use. The clause-by-clause reading is in EU AI Act Article 26: deployer obligations when you run an AI agent.

What are the steps, and what evidence does each leave?

1. Legal: confirms your role per agent and capability. Evidence: the role determination.
2. Compliance: confirms which capabilities are high-risk. Evidence: the classification record.
3. Agent owner: obtains and files the provider's instructions for use and maps each to how the agent is configured. Evidence: the instructions and the mapping.
4. Agent owner: names oversight people for each high-risk capability, with authority to reject, stop or lower depth. Evidence: the named approver roster and their authority.
5. Team leaders: ensure those people are competent, through the literacy programme set out in meeting AI literacy duties for staff who supervise AI agents. Evidence: training and assessment records.
6. Agent owner: controls input data: which sources the agent may use, their provenance and their fitness for the purpose. Evidence: the sources list and its review.
7. Agent operations: monitors operation: sampling, drift checks, held queue, refusals. Evidence: weekly reviews and decisions.
8. Agent operations: keeps logs of every verdict, append-only, for the required period. Evidence: the log store and reconstruction tests.
9. Compliance: sets up incident reporting, including what counts as serious and who reports. Evidence: the procedure and any reports made.
10. Compliance and HR: informs people: customers through notices and disclosures, workers or their representatives before use in the workplace. Evidence: notices and records of consultation.
11. Compliance: runs the fundamental rights impact assessment where required. Evidence: the assessment.

A one-page version is the deployer obligations checklist under the EU AI Act.

What does each obligation mean when the system acts?

Oversight must be able to stop an action, not just review it afterwards. For an agent, that means held actions for decisions that matter and the ability to lower depth at once.

Input data includes everything the agent reads. Not just training data, which you do not control, but the documents, records and messages it uses at the moment of action. Provenance labels show which inputs were verified.

Monitoring is continuous. An agent's behaviour can change with its sources or the cases it meets. Weekly review and drift checks are the practical form.

Suspension must be possible. If a risk appears, you must be able to stop use. Runbooks for lowering depth and taking work back are how you do it.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

What are the signs it is going wrong?

Oversight is a name on a list. The named person has no time, no training and no authority to reject.

Instructions for use were never read. The provider's limits are not reflected in the agent's bounds.

Logs are the provider's problem. The deployer assumes the provider keeps them, and cannot produce them when asked.

Workers learn about the agent from the agent. Consultation and information happen after go-live, if at all.