The EU AI Act and AI agents: what changed in 2026

The short answer: as of 2 August 2026 the AI Act's transparency obligations apply and the European AI Office can enforce against general-purpose AI model providers. The headline high-risk obligations — risk management, data governance, logging, human oversight, cybersecurity and post-market monitoring — were rescheduled: December 2027 for Annex III systems and August 2028 for Annex I product systems. If you deploy agents in the EU, the work to do now is disclosure, record-keeping and knowing which of your agents would land in scope later.

Third-party figures on this page were checked on 26 September 2026 against the publishing organisation's own release or report. Research in this field ages quickly and definitions vary between surveys — treat any undated statistic with suspicion, including on our own pages.

Why agents are treated differently from chatbots

The Act's obligations are written around what a system does. A chatbot produces text, and the duties attached to it are largely about making clear that a person is talking to a machine. An agent takes multi-step actions in other systems — it sends, books, files, updates a record, calls an API. Every duty that applies to output therefore has to cover actions too: the logging has to show what was done, the human oversight has to be able to intervene before it is done, and the traceability has to survive the fact that the action happened in someone else's system.

This is also why multi-agent setups need a moment's thought. If one agent delegates to another, the compliance boundary follows the function, not the org chart — a specialist agent performing a regulated step is in scope whether it was invoked by a person or by another agent. Our explainer on agent-to-agent delegation covers how those hand-offs are recorded.

What is in force now

• Transparency duties. People interacting with an AI system should be able to tell that they are. Synthetic content carries a disclosure obligation. For an agent answering on WhatsApp or email, this is a copy and configuration question more than an engineering one — and it is the single most common gap we see.
• Prohibited practices and AI literacy have applied since February 2025. The literacy duty is soft-edged and easy to ignore, and it is the one an auditor can check cheaply: can the people operating this agent explain what it does and where it stops?
• General-purpose AI model obligations have applied to providers since August 2025; from August 2026 the AI Office can enforce them. If you use a commercial model through an API you are a deployer, not a provider — but the provider's documentation is what you will lean on for your own records, so it is worth knowing what they publish.

What moved, and why you should not treat it as a reprieve

The high-risk obligations now land in December 2027 (Annex III — the use-case list covering employment, credit, essential services, education and similar) and August 2028 (Annex I — AI embedded in regulated products). Two reasons not to bank the delay:

1. The controls are the ones a competent deployment needs anyway. Logging what an agent did, keeping a human in the loop on consequential actions, monitoring behaviour after launch — these are the same controls that stop an agent project being cancelled. You are not building them for Brussels.
2. Retrofitting traceability is expensive. A system that never recorded who approved what cannot be made to have recorded it. Every month an agent runs without an audit trail is a month of history you cannot reconstruct.

A practical readiness list for agents

Ordered by how much they cost to add later:

1. An action log, not a chat log. What the agent did, in which system, at whose instruction, with what result. A transcript of the conversation is not the same record.
2. A human-oversight point on consequential actions. Not an all-or-nothing switch: a boundary around the irreversible, with everything upstream running unattended. See approvals and trust.
3. An inventory. Which agents exist, what each one is connected to, which would be high-risk if the use case shifted. Most organisations discover at this step that they have more agents than they thought.
4. Disclosure in the channels people actually use. If the agent answers on a messenger, the disclosure belongs there, not in a policy page.
5. A data-handling answer. What the agent reads, where that lands, and who can retrieve it. This is where the AI Act conversation meets the residency conversation — see where your agents actually run.
6. Named accountability. One person who can say what the agent is for and where it stops.

Provider or deployer?

Most businesses reading this are deployers: you use a model or a platform someone else provides, and you put it to work. Deployer duties are lighter than provider duties but they are not nothing — they centre on using the system as intended, maintaining human oversight, keeping logs within your control, and informing the people affected. You can become a provider by accident, most commonly by putting your own name on a system, or by modifying one substantially enough that you are effectively shipping something new.

If you are outside the EU

The Act reaches systems whose output is used in the EU. A Singapore or US company whose agent answers EU customers is inside the perimeter for those interactions. In practice most teams end up running one standard rather than two, because maintaining an agent that behaves differently by customer geography is more expensive than the compliance it avoids.

The same logic applies in reverse to the American picture, where the constraint is a patchwork of state statutes rather than one regulation — covered in the US state AI law patchwork.

This is an orientation, not legal advice. Deadlines and interpretations in this area have moved more than once; confirm current obligations with a qualified adviser before relying on them.