EU AI Act hits agentic systems
Europe’s new AI law is becoming harder to follow for systems that plan and act on their own rather than just return predictions. The latest analysis warns that “agentic” AIs will force organisations to classify use cases, document oversight and manage risks across automated workflows — and practitioners find mapping real products to the law’s categories surprisingly difficult (artificialintelligence-news.com) (dev.to). The Commission is also expanding the rulebook to include energy and emissions measurements for AI, signalling compliance will cover sustainability as well as safety (dig.watch).
A chatbot that only answers questions is one thing. A system that can decide, click, book, escalate, and trigger other software starts to look less like a calculator and more like a junior employee, and Europe’s Artificial Intelligence Act was mostly written around categories that are easier to spot than that. (digital-strategy.ec.europa.eu) The law is already turning on in stages. Rules for banned uses started applying on 2 February 2025, rules for general-purpose artificial intelligence models started on 2 August 2025, and most rules for high-risk systems apply from 2 August 2026. (digital-strategy.ec.europa.eu 1) (digital-strategy.ec.europa.eu 2) (artificialintelligenceact.eu) That phased rollout works fine when a product fits one box. It gets messy when one product contains a general-purpose model, a decision engine, a workflow tool, and a human approval screen all in the same chain. (dev.to) The Act is built on risk tiers. Some uses are banned outright, some are labeled high-risk and come with testing, logging, documentation, and human-oversight duties, and some systems mainly face transparency rules such as telling people they are dealing with artificial intelligence or seeing synthetic content. (digital-strategy.ec.europa.eu 1) (digital-strategy.ec.europa.eu 2) An agentic system scrambles those tiers because the model is not the whole product. The legal question shifts from “what model is this” to “what job is this system performing at each step, and who is responsible when it acts.” (artificialintelligence-news.com) (dev.to) That is why companies are being pushed toward use-case mapping. If one agent drafts a loan summary, another checks identity documents, and a third sends an approval recommendation, the compliance team may have to classify each function separately instead of treating the stack as one harmless assistant. (dev.to) (digital-strategy.ec.europa.eu) The European Commission has been filling in the blanks with guidance. It published guidelines on the definition of an artificial intelligence system on 6 February 2025 and guidelines on prohibited practices on 4 February 2025, which shows how much interpretation is needed before many teams can even decide which rulebook applies. (digital-strategy.ec.europa.eu 1) (digital-strategy.ec.europa.eu 2) The same pattern is showing up around transparency. Article 50 transparency duties come fully into effect on 2 August 2026, and European Union materials released in March 2026 started translating those duties into practical marking and labeling expectations for artificial intelligence-generated content. (rpclegal.com) (www.nicfab.eu) Now the compliance load is spreading beyond safety and rights into energy use. The Commission opened a targeted consultation in April 2026 on measuring energy consumption and emissions of artificial intelligence models and systems, and it said the work could support a future energy and emissions label. (digital-strategy.ec.europa.eu) So the paperwork around an autonomous system is starting to look like an audit trail for a factory line. Teams may need records for what the system was designed to do, where a human can intervene, what risks each step creates, and how much electricity the whole setup burns while doing it. (artificialintelligence-news.com) (digital-strategy.ec.europa.eu) Europe is not banning agents as a category. It is forcing builders to break the magic trick into parts, assign each part a legal role, and prove that a system acting across multiple tools still has a named human or company on the hook when something goes wrong. (digital-strategy.ec.europa.eu) (dev.to)
