Article 26 is the part of the EU AI Act written for the business that uses a high-risk AI system rather than the one that built it. If you run a high-risk AI agent, you must use it as instructed, assign trained people to oversee it, control its inputs, monitor it, keep its logs and tell the people it affects.
What does Article 26 require, duty by duty?
In substance, a deployer of a high-risk system must:
- Use it as instructed. Take appropriate technical and organisational measures to use the system in line with the provider's instructions for use.
- Assign human oversight. Give oversight to people with the competence, training and authority to do it, and the support they need.
- Control input data. Where the deployer controls the inputs, make sure they are relevant and sufficiently representative for the intended purpose.
- Monitor and report. Monitor operation on the basis of the instructions; inform the provider, and where needed the authorities, of risks, and suspend use if the system presents a risk; report serious incidents.
- Keep logs. Retain the automatically generated logs under its control for an appropriate period, with a statutory minimum unless other law, especially data protection law, says otherwise.
- Inform workers. Before putting a high-risk system into use at work, tell workers' representatives and affected workers.
- Register, if a public body. Public authorities must register their use in the EU database.
- Use the provider's information in the DPIA. Where a data protection impact assessment is needed, draw on the provider's information to complete it.
- Inform affected people. Tell natural persons when a listed high-risk system makes or helps make decisions about them.
- Cooperate with authorities.
Financial institutions can meet the monitoring and log duties within the governance and record-keeping they already run under financial services law.
When are you the deployer?
You are a deployer when you use an AI system under your own authority in a professional activity. Most firms that buy an agent platform and configure agents on it are deployers. You can become a provider of a high-risk system if you put your name on it, change its intended purpose so that it becomes high-risk, or modify it substantially. The distinction is unpacked in provider or deployer: your role under the EU AI Act and, more briefly, in what a deployer is under the EU AI Act.
For agents, the line is less clean than for bought models. Writing an agent's brief, choosing its sources and connecting it to a payment system are configuration choices that shape behaviour heavily. They usually remain deployment, but a change that repurposes an agent into a listed high-risk use is the kind of step that can tip a deployer into provider territory.
What do these duties mean for an agent that acts?
Three duties change shape.
Oversight for a scoring model means a person reviews scores. For an agent it means deciding, per capability, which actions a person must approve before they happen. Oversight that only reviews logs after the event cannot stop a payment. The competence requirement also bites: the overseer must know what the agent can do and what it cannot see.
Input data for an agent means its runtime sources: the records, documents and inboxes it reads. Article 26 makes the deployer answerable for their relevance and representativeness, which in practice means a source register and a rule on untrusted content.
Monitoring means watching actions, not just outputs. Which actions were refused, which were held, how long holds waited, whether overseers ever overruled the agent. Our piece on Article 26 monitoring duties goes further into the measures.
How do you meet Article 26 for an AI agent?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Use per instructions | Capabilities enabled only within the provider's stated purpose | Configuration record mapped to the instructions version |
| Competent human oversight | Named overseer per capability, trained on that agent; hold points set by risk | Assignment list, training records, approval records |
| Relevant input data | Source register with owner and trust class for each source | Register and retrieval records per action |
| Monitor and suspend | Dashboards on refusals, holds, overrides and errors; a tested stop control | Monitoring reports; suspension log |
| Report risks and incidents | Escalation path to provider and authority | Incident tickets with notification dates |
| Keep logs | Append-only action record retained on a set schedule | Retention schedule and integrity checks |
| Inform workers | Consultation before workplace use | Notices and representative minutes |
| Inform affected people | Notice in decision letters and channels | Templates and sent copies |
The checklist version, with sign-off fields, is in our deployer obligations checklist under the EU AI Act.
What will a supervisor or auditor ask?
- Is this agent high-risk, and what makes you the deployer rather than the provider?
- Who oversees it? Show their training and their approvals.
- Which instructions for use applied, and how does your configuration stay within them?
- What did monitoring find last quarter, and what did you do?
- Have you ever suspended it? How fast could you?
- How were affected customers and staff told?
What does Article 26 not cover?
It covers high-risk systems only, apart from duties set elsewhere for all systems, such as AI literacy and transparency to people. It does not make the deployer answerable for the provider's design duties, such as building logging in or declaring accuracy. It does not include the fundamental rights impact assessment, which is Article 27 and applies to certain deployers only. And it does not replace sector rules: a bank's conduct, model risk and outsourcing obligations still apply in full. Our guide to AI agent regulation, mapped to controls shows where each sits.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Does Article 26 apply to an agent used only by staff?
Yes, if the agent is high-risk. Internal use is still deployment. Some listed uses, such as decisions about recruitment or promotion, are internal by nature, and those also trigger the duty to inform workers.
Can oversight be shared across many agents?
Yes, provided each overseer has the competence, authority and time for what they oversee. One person approving every action of ten agents with no time to read them is not oversight a supervisor will accept.
What counts as suspending an agent?
Stopping it from taking actions while the risk is assessed, and handing work back to people. Design the stop so it can be used quickly by the overseer, and test it before you need it.