EU AI Act already covers agentic AI, standards advisers say
The European Commission says agentic AI falls within the EU AI Act. Standards Norway advisers argue that binding risk controls matter more than voluntary tech-led governance.
By Amara Osei
3 min read
What's News
- The European Commission says agentic AI falls within the EU AI Act because it meets the definition of an AI system.
- Agentic AI often uses large language models, which are regulated as general-purpose AI models.
- The essay identifies four high-risk categories: recruitment, essential services, medical devices and safety-critical infrastructure.
- The White House Accord on Superintelligence is a voluntary pact involving X, Meta, Anthropic, OpenAI and Google.
- Dr. Muller works on a committee nearing completion of harmonized risk-management standard 18228.
By Amara Osei
The European Commission says agentic AI falls within the EU AI Act because it meets the broader definition of an AI system.
Standards Norway committee members writing in a Unite.AI essay reject claims that the law leaves AI agents outside its reach. They say arguments repeated by Morgenbladet, Dagensnæringsliv and University of Oslo academics rest on “fears, not on fact” and could weaken support for enforcement.
What does the AI Act cover?
The authors say they work closely with the AI Act as appointed members of Standards Norway’s national AI committee. They also sit on European committees developing harmonized standards for providers of high-risk AI systems, which help companies meet the law’s legal requirements.
The European Commission’s plain-language FAQs confirm that “agents” and “agentic AI” fall within the regulation. The law regulates what technology does and the risks created by those functions, rather than naming individual technologies.
Agentic systems can trigger several regulatory routes:
- Agentic AI often uses large language models, which fall under the rules for general-purpose AI models.
- When an agentic AI system interacts with humans, the AI Act’s transparency requirements apply.
- Models classified as posing systemic risk face additional provider obligations, including capability assessments covering autonomous and agentic features.
The absence of the word “agent” from the AI Act therefore does not create a regulatory gap. According to the essay, the omission reflects the law’s technology-neutral design.
Which uses carry high-risk obligations?
The AI Act focuses on risks that could affect health, safety and fundamental rights, a priority the authors say appears in the first sentence of Article 1. They also describe the law as market-oriented because it seeks to encourage AI use and support innovation.
The authors list four high-risk application areas:
- Recruitment systems.
- Allocation of essential services.
- Medical devices.
- AI systems integrated into critical infrastructure where safety is paramount.
Providers whose systems could affect fundamental rights must establish comprehensive risk-management systems. Those systems must account for how users might misuse the AI, according to the Standards Norway committee members.
The authors also answer critics who consider perfect risk removal impossible. They argue that no international regulatory framework can predict every future risk, making targeted controls more realistic than an assumption of zero harm.
Why do self-regulation claims matter?
Anthropic, OpenAI and Meta, along with some participants in the Norwegian debate, have portrayed AI and agentic AI as too complex, advanced or dangerous for the EU to regulate, the essay says. The authors counter that these companies have lobbied for less stringent implementation and now favor voluntary controls.
They describe that approach as regulatory capture: a process in which entities subject to regulation increasingly shape the rules for their own benefit. Their concern centers less on voluntary standards themselves than on replacing enforceable obligations with promises from the companies being regulated.
The authors cite the White House Accord on Superintelligence as a US example. It is a “voluntary pact” involving leaders from X, Meta, Anthropic, OpenAI and Google, under which the companies promise to regulate themselves. The accord says: “Over time, it may make sense to codify these steps into laws or regulations.”
Europe’s framework differs because the EU AI Act is binding law. The authors warn that regulatory capture can still take hold if technology companies undermine public confidence in that law. If major providers persuade people that the AI Act cannot handle their systems, support for enforcement could fade, and reversing that shift would be difficult.
The authors acknowledge that some AI systems may produce consequences they cannot foresee. They cite Dr. Muller’s work on a committee nearly finished developing the harmonized risk-management standard identified as 18228 as the practical response.
They concede that the AI Act is not perfect and will probably need further development. Their warning is equally direct: “portraying agentic AI as an area outside of the EU AI Act only builds uncertainty, not safety.”
As harmonized standards near completion, the authors argue that credible enforcement will determine whether agentic AI remains inside Europe’s binding regime or gives way to voluntary promises from technology providers.
Source: unite.ai
More from Amara Osei
Show full bio
Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.
2,358 articles