Policy & Regulation

EU AI Act Already Covers Agentic AI, Standards Body Members Say

Members of Norway's national AI committee say the Commission's own FAQs confirm agentic AI meets the AI Act's definition of an AI system — and warn that "regulatory gap" claims pave the way for regulatory capture.

Yes, the EU AI Act Covers Agentic AI—Here’s How
Yes, the EU AI Act Covers Agentic AI—Here’s Howpixel2013 / pixabay
By Marcus Bennett5 min read

Why it matters

  • The European Commission's AI Act Service Desk FAQs confirm agentic AI falls within the AI Act's definition of an AI system.
  • Systemic-risk model providers must assess capabilities including autonomous and agentic features.
  • The White House Accord on Superintelligence is a voluntary self-regulation pact with X, Meta, Anthropic, OpenAI, and Google.
  • The harmonized risk-management standard, numbered 18228, is nearly complete.
  • High-risk domains include recruitment, essential services, medical devices, and critical infrastructure.

The European Commission has stated plainly that agentic AI falls within the scope of the EU AI Act, because it meets the regulation's definition of an AI system. That clarification, published in plain-language FAQs by the Commission's AI Act Service Desk, sits at the center of a growing argument over whether autonomous AI agents escape European oversight — an argument that members of Norway's national AI standards committee now say is not only wrong but dangerous.

The rebuttal comes from appointed members of Standards Norway's national AI committee who also sit on the European committees developing harmonized standards — technical specifications that help companies meet the AI Act's legal requirements — for providers of high-risk AI systems. Their position is blunt. Agentic AI is regulated under the EU AI Act today, and claims to the contrary erode confidence in a binding law.

The "regulatory gap" narrative has spread widely. It has appeared in reputable Norwegian outlets Morgenbladet and Dagens Næringsliv, and has been repeated by leading academics at the University of Oslo. The committee members identify two fundamental flaws in these claims, and argue they are grounded in fears rather than fact.

What does the AI Act actually regulate?

The AI Act never uses the word "agent." That omission is deliberate, not a loophole. The regulation does not target specific technologies by name. It regulates what a technology is designed to do and the risks those functions create.

The Commission's Service Desk FAQs on "agents" and "agentic AI" confirm the point directly: such systems fall within the regulation's scope because they meet the broader definition of an AI system. Several concrete mechanisms apply:

  • Agentic AI is frequently built on large language models, which the AI Act regulates as general-purpose AI models.
  • When an agentic system interacts with humans, the regulation's transparency requirements are triggered.
  • When a model is classified as posing a systemic risk (a category applying to the most powerful general-purpose models), additional obligations apply to the provider — including assessments of the model's capabilities, explicitly covering its autonomous and agentic features.

That last point matters for the debate. The law does not merely tolerate agentic behavior; it requires providers of systemically risky models to evaluate it.

The stakes are structural. The AI Act's first article, first sentence, states that the regulation addresses risks posed by AI systems to health, safety, and fundamental rights. The same framework is also market-oriented: the law is designed to encourage AI use and support innovation. In the authors' reading, the AI Act is pro-AI, regulating only certain applications and domains within that pro-innovation context.

Which uses are pre-classified as high-risk?

The Act does not wait for harm to occur before imposing duties. It pre-classifies entire domains as high-risk, including AI systems used in:

  • recruitment;
  • the allocation of essential services;
  • medical devices;
  • systems integrated into critical infrastructure where safety is paramount.

Providers of AI systems — agentic or otherwise — that could affect fundamental rights must establish comprehensive risk management systems. Those systems must include assessments of how users may misuse the AI system. The duty to anticipate misuse, not merely to react to it, is embedded in the law.

Critics who want even stricter rules often point to the structural impossibility of removing every risk in advance. The authors concede the point but reject the conclusion. No international regulatory framework could have anticipated every future risk. The AI Act's answer to unforeseeable consequences is procedural: mandatory risk management, not exhaustive prediction.

Dr. Muller, one of the authors, works on the committee that is nearly finished developing the harmonized standard for risk management, numbered 18228. That standard, once complete, will give providers a concrete technical path to compliance.

Why do the authors invoke regulatory capture?

The sharpest charge in the piece concerns who benefits from the "too complex to regulate" narrative. The authors point to Anthropic, OpenAI, and Meta, along with certain voices in the Norwegian debate, as promoting the impression that agentic AI is too advanced for the AI Act, too advanced for the EU, and too dangerous to be regulated by anyone other than the companies themselves.

The authors urge caution. These are the world's most powerful technology companies, and they have lobbied hard for a less stringent implementation of the AI Act. Their proposed solution is to regulate themselves.

The authors label this a classic mechanism of regulatory capture — a process in which laws and regulations are gradually shaped by the very entities they are intended to govern, for those entities' own benefit.

They cite a concrete precedent: the White House Accord on Superintelligence, described in the source as a "voluntary pact" with leaders from X, Meta, Anthropic, OpenAI, and Google, published under the heading "Inaugurating The Era Of Super Intelligence." The document acknowledges that binding rules might eventually be warranted, stating: "Over Time, it may make sense to codify these steps into laws or regulations."

The authors' reading of that sentence is pointed. Over time. Perhaps. For now, Americans are expected to trust that major tech companies will regulate themselves. Europe, by contrast, has a binding law: the AI Act.

How could a strong law still be weakened?

Regulatory capture does not require repealing a statute, the authors argue. It can proceed through the erosion of public confidence in the law. If tech giants can convince people that the AI Act is not good enough, support for its enforcement will decline.

Claims that agentic AI is too advanced, too smart, or too dangerous for the AI Act serve exactly that function, in the authors' framing. Each repetition weakens public backing for regulation while implying that AI companies are better positioned to manage the risks of their own systems — the same argument that historically enables capture. Once regulatory capture takes hold, the authors warn, it is difficult to reverse.

The piece closes on a measured note rather than a defensive one. The EU AI Act is not perfect, the authors concede, and it will likely need further development. But portraying agentic AI as an area outside the law builds uncertainty, not safety. The forward path they identify is already underway in the standards bodies: the harmonized risk-management standard, developed under Dr. Muller's committee, is nearly finished — meaning providers of agentic systems will soon have a concrete European template for assessing exactly the autonomous capabilities the gap narrative claims no one knows how to govern.

Source: unite.ai

Share this article:

More from Marcus Bennett

Marcus Bennett

Show full bio

Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

791 articles

Related articles

  1. Over 1,000 State AI Bills in 2025, and No Rulebook to Reconcile Them
  2. Enterprise AI Guardrails Are Shifting From Frontier Fear to Buyer Contracts
  3. The EU's Next Chapter in AI Is Agentic, New Analysis Claims
  4. FTC Opens Industry-Wide Probe Into Anthropic, OpenAI Over AI Agent Risks
  5. Australian AI governance overhaul proposed in Bond Law Review

« Previous articleNext article »