- Home
- AI agents
- Governance and regulation
- EU AI Act and AI agents
Govern and secure · Governance and regulation
EU AI Act and AI agents: what applies, and when
The EU AI Act applies to AI agents by what they do: an agent that talks to people carries the Article 50 transparency duty since 2 Aug 2026, and one in an Annex III high-risk use carries the high-risk duties from 2 Dec 2027. The Digital Omnibus moved only the high-risk dates. This is general information rather than legal advice.
In this guide: 83 pages
Key takeaways
- Regulation (EU) 2026/1744 entered into force on 27 Jul 2026 and moved Annex III high-risk obligations from 2 Aug 2026 to 2 Dec 2027 and Annex I obligations from 2 Aug 2027 to 2 Aug 2028. Everything else stayed on schedule.
- Since 2 Aug 2026 an agent that interacts with people must make that clear; the Commission’s examples are "a chatbot, AI agent, and avatar". The grace period for marking AI-generated output on systems already on the market ends on 2 Dec 2026.
- The Commission’s draft classification guidelines (19 May 2026) assess agentic systems as a whole: a component doing only a narrow procedural or preparatory task may still be classified as high-risk. The draft’s final status was unverified on 2 Nov 2026.
- For high-risk uses the agent duties are concrete: automatic logging (Article 12), logs kept at least six months (Articles 19 and 26), competent human oversight (Articles 14 and 26), monitoring and suspension (Article 26). Build them now; the dates are fixed.
What changed with the Digital Omnibus, and what did not?
The Digital Omnibus on AI changed the high-risk dates, softened the AI literacy duty, added a prohibition and adjusted fines for small mid-caps; it did not touch the prohibitions already in force, the general-purpose model duties or the transparency duties. The table sets the two side by side.
Before and after Regulation (EU) 2026/1744, as the Commission’s page (updated 19 Oct 2026) and the Parliament’s adopted text show it.
| Obligation | Original date | After the Omnibus |
|---|---|---|
| Prohibited practices (Article 5) and AI literacy (Article 4) | 2 Feb 2025 | Unchanged; the literacy duty was softened |
| General-purpose AI model obligations | 2 Aug 2025 | Unchanged |
| Article 50 transparency (people told they deal with AI; marking of generated content) | 2 Aug 2026 | Unchanged; a grace period for marking on systems already on the market runs to 2 Dec 2026 |
| New prohibition: non-consensual intimate imagery and child sexual abuse material | Not in the original Act | Applies from 2 Dec 2026 |
| High-risk obligations, Annex III uses (employment, credit, education, essential services and others) | 2 Aug 2026 | 2 Dec 2027 |
| High-risk obligations, Annex I product-embedded systems | 2 Aug 2027 | 2 Aug 2028 |
| Fines for transparency breaches | Up to 15 million euros or 3% of worldwide annual turnover | Unchanged; lighter caps for small mid-caps added in Article 99(6a) |
The legislative trail: Commission proposal 19 Nov 2025; Council mandate 13 Mar 2026; Parliament position 26 Mar 2026; provisional agreement 7 May 2026; Parliament first reading 16 Jun 2026; Council adoption 29 Jun 2026; signature 8 Jul 2026; Official Journal 24 Jul 2026; in force 27 Jul 2026. EUR-Lex blocked automated reads during our research, so quote article wording only after checking it there; a corrigendum dated 29 Sep 2026 is listed and was not examined.
Which tier does an agent fall into?
An agent falls into the tier of the decision or interaction it is part of: prohibited if it does a banned thing, a high-risk candidate if it forms part of an Annex III or Annex I use, transparency-only if it interacts with people or generates content, and minimal otherwise. The agent’s model, framework and platform do not change the tier; the classification conditions and the draft guidelines do, so assess the whole system with counsel.
The transparency tier catches most enterprise agents today. Any agent that chats with customers or staff must make clear that it is an AI system, and any agent that produces images, audio or video must mark the output. The Commission’s own list of examples for Article 50 reads "a chatbot, AI agent, and avatar", so an internal HR assistant, a customer-service voice agent and a drafting agent that produces media all sit here from 2 Aug 2026.
The high-risk tier is where agent design matters most. The Commission’s draft guidelines of 19 May 2026 say that "complex systems made up of several AI components, including so-called agentic AI systems, must be assessed holistically", and that even a component performing only a "narrow procedural or preparatory task" may still be classified as high-risk. A screening agent that ranks candidates before a recruiter decides, or a document agent that prepares a credit file, may therefore be in scope even though a person signs the outcome. The draft was open for consultation until 23 Jun 2026 and its final status was unverified on 2 Nov 2026.
The prohibited tier applies to agents too. The practices banned since 2 Feb 2025 do not stop being banned because an agent rather than a person performs them, and a new prohibition on non-consensual intimate imagery and child sexual abuse material applies from 2 Dec 2026. An agent with image generation tools needs a guardrail for it before that date.
What do the provider and deployer duties mean for an agent team?
For a high-risk use, the provider must build an agent that logs automatically and can be overseen, and the deployer must assign competent people, keep the logs, monitor the agent and suspend it when a risk appears; most enterprises are deployers, and a company that builds its own agent can be both. The table turns each article into an engineering requirement.
High-risk duties that bite for agents from 2 Dec 2027 (Annex III) and 2 Aug 2028 (Annex I), with the article and the build requirement.
| Article | Who | What it asks | What to build |
|---|---|---|---|
| Article 12, record-keeping | Provider | The system must technically allow automatic recording of events over its lifetime | A trace per run with run id, timestamps, model and prompt version, tool calls with arguments and results, retrieved sources, output and errors |
| Article 14, human oversight | Provider | The system must be designed so that people can oversee it, understand its limits and intervene or stop it | An approval step before irreversible actions, a stop that ends the run, and a user interface that shows what the agent did |
| Article 19(1), logs | Provider | Keep automatically generated logs under your control for at least six months unless other law says otherwise | Retention of at least six months on the trace store, with legal hold |
| Article 26(2), oversight by the deployer | Deployer | Assign oversight to people with the necessary competence, training and authority | Named approvers per agent, trained, with authority recorded; approval events in the trace |
| Article 26(5), monitoring | Deployer | Monitor operation and suspend use if a risk appears | Runtime alerts, a tested kill switch and suspension events logged |
| Article 26(6), deployer logs | Deployer | Keep the logs the system generates for at least six months | The same retention, under the deployer’s control, even when the vendor hosts the agent |
| Article 50, transparency | Provider and deployer | People are told they interact with AI; generated content is marked | A disclosure in every conversational agent and output marking for media, live since 2 Aug 2026 |
One consequence is often missed: Article 26(6) puts the six-month log duty on the deployer, so a company that buys a hosted agent still needs the vendor to export or retain the logs on its behalf, in a contract, before 2 Dec 2027.
What should you do before each date?
The work orders itself by date: the transparency duties are live, the next prohibition and the marking grace period end on 2 Dec 2026, and the high-risk dates in 2027 and 2028 are fixed. We recommend this order.
- Now: add the AI disclosure to every agent that talks to people and mark generated media. Article 50 has applied since 2 Aug 2026 and the AI Office has had enforcement powers since the same day.
- Also now: classify each agent by the decision it influences, with counsel, using the holistic reading for multi-agent systems. Record the tier in the registry.
- By 2 Dec 2026: add a guardrail against the newly prohibited content to any agent with image tools, and finish output marking on systems that were already on the market.
- Ahead of 2 Dec 2027: for Annex III candidates, turn on per-run tracing with tool-call arguments, set retention to at least six months (longer if another rule applies), name trained approvers, test the stop, and get the vendor’s log export written into the contract.
- For 2 Aug 2028: repeat for agents embedded in Annex I regulated products, where the product regime adds its own conformity steps.
- At each date: re-read the Commission’s AI Act page and EUR-Lex, since the corrigendum and the final classification guidelines may change wording.
How do Microsoft, AWS and Google implement it?
What each cloud offered on 2 Nov 2026 to help with the EU AI Act duties above.
Capability, Microsoft, AWS, Google
| Capability | Microsoft | AWS | |
|---|---|---|---|
| Mapping templates | Purview Compliance Manager premium template for the EU Artificial Intelligence Act | No EU AI Act framework; Audit Manager is in maintenance mode from 30 Apr 2026 with no new frameworks | Not verified |
| Article 12 logging | Foundry tracing (generally available for prompt and hosted agents); Purview unified audit log for Copilot and Copilot Studio | AgentCore Observability traces on CloudWatch; CloudTrail for AgentCore API calls | Agent Observability (generally available 18 Jun 2026) |
| Article 14 and 26 oversight | Copilot Studio multistage and AI approvals (preview); Agent 365 block and quarantine actions | AgentCore Policy to deny tool calls (generally available 3 Mar 2026); Amazon A2I is closed to new customers | Agent Gateway policies; ADK human-input steps in graph workflows |
| Retention control | Purview retention policies on AI interactions; Application Insights retention | CloudWatch log retention settings; S3 lifecycle for exported traces | Cloud Logging retention settings |
None of these is a compliance claim by the vendor or by QueryNow; each is a feature that can hold the evidence the articles describe.
Frequently asked questions
Does the EU AI Act mention AI agents?
The Act regulates by use rather than by technology, so the text does not define agents. The Commission’s guidance does name them: its 2 Aug 2026 explanation of Article 50 lists "a chatbot, AI agent, and avatar" as examples, and its draft classification guidelines of 19 May 2026 say agentic systems must be assessed holistically.
Are the high-risk obligations postponed again?
They were postponed once, by Regulation (EU) 2026/1744: Annex III to 2 Dec 2027 and Annex I to 2 Aug 2028. The Commission’s original proposal had tied the dates to the availability of standards; the final text uses fixed dates. On 2 Nov 2026 no further postponement was proposed.
Is an internal agent for employees covered?
Yes if it interacts with people, which triggers the Article 50 disclosure since 2 Aug 2026, and possibly if it forms part of an Annex III use such as hiring, promotion or task allocation, which can trigger the high-risk duties from 2 Dec 2027 once the whole system is assessed. An agent that only summarises documents for an analyst, with no interaction and no such decision, sits in the minimal tier.
How long must we keep agent logs under the EU AI Act?
At least six months for high-risk systems, under Article 19(1) for providers and Article 26(6) for deployers, unless applicable EU or national law provides otherwise, including data-protection law. Sector rules, Colorado’s three-year record rule for US operations and your own policies may extend it, so set the trace store to the longest period that applies.
Is this page legal advice?
No. It is general information with dates and article numbers as the European Commission, Parliament and Council pages showed them on 2 Nov 2026. Article wording is quoted from an unofficial reading aid where EUR-Lex blocked automated access. Confirm every obligation with counsel before you rely on it.
More in Governance and regulation
Sources
Each source was read on 2 November 2026. The date shown is the date on the page, or the page’s ms.date for Microsoft Learn; an undated page shows the date it was read.
- AI Act: regulatory framework for AI
- Safer and more transparent AI
- Regulation (EU) 2026/1744, Digital Omnibus on AI (Official Journal)
- European Parliament position P10_TA(2026)0198 on the Digital Omnibus on AI
- Council agrees position to streamline rules on artificial intelligence
- Artificial intelligence: Council gives final green light to simplify and streamline rules
- EU Digital Omnibus on AI enters into force
- EU AI Act Omnibus agreement: postponed high-risk deadlines and other key changes
- CSA research note: EU AI Act high-risk deadline and the Omnibus, "Deferred, Not Cancelled"
- EU AI Act update: the European Commission publishes draft guidelines on high-risk AI systems
- EU AI Act Article 12, record-keeping (reading aid)
- EU AI Act Article 19, automatically generated logs (reading aid)
- EU AI Act Article 26, obligations of deployers (reading aid)
- EU AI Act Article 99, penalties (reading aid)
- Regulations in Compliance Manager
- Microsoft Purview data security and compliance protections for generative AI apps
- Set up tracing for agents in Microsoft Foundry
- AWS Audit Manager availability change
- Policy for Amazon Bedrock AgentCore generally available
- Gemini Enterprise Agent Platform release notes
Planning an AI program?
Talk to the team that builds it and runs it after launch.
- Built to SOC 2, HIPAA and GDPR standards
- EU AI Act-aligned delivery
- Microsoft Solutions Partner