SYNTROX.ai
Blog

EU AI Act Article 26: What Companies Deploying AI Agents Have to Do

6 min read

If you use a high-risk AI system, Article 26 makes you answerable for how it runs: use it as the provider's instructions say, assign trained people with authority to oversee it, keep the input data you control relevant, monitor it and suspend it if it presents a risk, keep its logs for at least six months, and tell workers and affected people. For Annex III systems this applies from 2 December 2027.

On this page

Most organisations running AI agents did not build the model underneath. They configured it, connected it to their systems and put it to work. In the EU AI Act's vocabulary that makes them deployers, and Article 26 lists what deployers of high-risk systems owe.

This is a practical reading for teams running agents, not legal advice.

Who is a deployer under the EU AI Act?

Article 3(4) defines a deployer as anyone "using an AI system under its authority", other than in a personal, non-professional activity. If your company runs an agent in its own operations, you are very likely its deployer.

You can also become its provider. Article 25 treats a deployer as a provider, with the provider's much heavier obligations, if it puts its name or trademark on a high-risk system, makes a substantial modification to one, or changes the intended purpose of a system, including a general-purpose one, so that it becomes high-risk.

That last case matters for agents. Suppose you build an agent on a general-purpose model to filter job applications, a use listed in Annex III, point 4(a). The provider role may now be yours. Check with counsel before assuming you are only a deployer.

Whether an agent is high-risk at all depends on what it decides; we walk through that test in our guide to agent audit trails under Article 12.

When do Article 26 obligations apply?

The Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026, moved Article 26's start date. Under the amended Article 113:

  • Annex III systems (employment, credit, essential services and the other listed areas) are covered from 2 December 2027.
  • Annex I systems (AI in products covered by EU product-safety law) are covered from 2 August 2028.

One deployer duty already applies. Article 4, on AI literacy, has applied since 2 February 2025 to providers and deployers of all AI systems, high-risk or not. As amended by the Omnibus, it asks them to take measures to support their staff's AI literacy, without guaranteeing any particular level for any individual.

EU AI Act deployer obligations: Article 26 for AI agents

Article 26 has twelve paragraphs. Set aside a saving clause, the general duty to cooperate with authorities, and the paragraphs that only concern public authorities and post-remote biometric identification, and nine duties remain for a company running an agent.

Article 26The dutyWhat it means for an agent
26(1)Use the system in accordance with the provider's instructions for useKeep the agent's tools, data and tasks inside the intended purpose those instructions describe
26(2)Assign human oversight to people with the necessary competence, training, authority and supportA named owner per agent who can understand it, override it and stop it
26(4)Where you control the input data, ensure it is relevant and sufficiently representativeGovern what the agent retrieves and what its tools return, not only the prompt
26(5)Monitor operation on the basis of the instructions for useWatch what the agent does in production, action by action
26(5)If you have reason to consider it presents a risk, inform the provider and the market surveillance authority, and suspend use; report serious incidentsA stop that works, and a reporting path written down in advance
26(6)Keep the logs under your control for at least six monthsA per-action record, retained and protected
26(7)Before using it at the workplace, inform workers' representatives and affected workersTell staff before an agent allocates or evaluates their work
26(9)Use the provider's Article 13 information for your data protection impact assessmentA DPIA that describes the agent as deployed, with its tools and data
26(11)Tell people they are subject to an Annex III system that makes or assists decisions about themDisclosure to candidates, applicants and customers

Breaching Article 26 falls under Article 99(4): fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher.

Three of those rows are harder for agents than the text suggests.

Input data is whatever the agent reads

Article 3(33) defines input data as data "provided to or directly acquired by" the system. An agent acquires data all the time: documents it retrieves, records it looks up, results its tools return. On our reading, Article 26(4) reaches all of that wherever you control the source. A superseded policy still in the retrieval index is an input data problem.

Oversight needs someone who can actually stop it

Article 26(2) is one sentence, but it leans on Article 14(4), which describes what the people assigned to oversight must be enabled to do: understand the system's capacities and limitations, stay aware of automation bias, interpret its output, decide not to use it or override it, and interrupt it through "a 'stop' button or a similar procedure" that brings it to a halt in a safe state.

An owner who cannot see individual actions, or whose only stop is an instruction to the agent, does not fit that description. See why approval steps often fail as oversight and why the stop has to be held outside the agent.

Monitoring has to be good enough to trigger a suspension

Article 26(5) asks you to act, not just to watch. If you have reason to consider that the system presents a risk to health, safety or fundamental rights (the Article 79(1) test), you inform the provider and the market surveillance authority "without undue delay" and suspend use. If you identify a serious incident, you inform the provider first, then the importer or distributor and the authorities. If you cannot reach the provider, the reporting rules of Article 73 apply to you, with a general outer limit of 15 days from becoming aware and shorter limits for the gravest cases.

That chain only works if your monitoring can tell a risk from noise, action by action, soon enough to matter. Financial institutions get one simplification: meeting their internal governance rules under financial services law counts as meeting the monitoring duty.

What deployers owe workers and the people affected

Two paragraphs are about telling people.

Workers. Under Article 26(7), an employer must inform workers' representatives and the affected workers before putting a high-risk system into service at the workplace. Suppose an agent allocates tasks or evaluates performance, both uses listed in Annex III point 4(b). The notice comes before the rollout, not after the pilot.

People decisions are made about. Under Article 26(11), deployers of Annex III systems that make or assist decisions about people must tell those people they are subject to the system.

Some deployers owe more. Article 27 requires a fundamental rights impact assessment before first use from public bodies, private entities providing public services, and deployers of credit-scoring and of life and health insurance pricing systems. It must describe, among other things, the oversight measures and what you will do if the risks materialise.

An Article 26 readiness checklist for AI agents

  1. Is each agent high-risk, and are you its deployer, or its provider under Article 25?
  2. Does the agent's configuration stay inside the provider's instructions for use?
  3. Who is the named oversight owner, and can they see, override and stop individual actions?
  4. Which data sources and tools feed the agent, and who keeps them current?
  5. What counts as "reason to consider" a risk, who decides, and how is use suspended?
  6. Are logs kept for at least six months, outside the agent's reach?
  7. Have workers and affected people been told, before use?

Where Syntrox fits

Syntrox sits outside the agent and records, for each trace it evaluates, the score, the confidence and the version of the judge that produced them: evidence for the monitoring Article 26(5) asks of deployers. Its judge, Nucleus, returns pass, fail or uncertain with a short reason that quotes the fragment that drove the verdict, which gives an oversight owner something specific to act on. The obligations stay with the deployer; what Syntrox adds is the evidence to meet them.