The EU AI Act and AI agents:
what it requires, and when.

The EU AI Act is the European Union's risk-based regulation of artificial intelligence. It does not single out agents; it classifies AI systems by how risky their use is and attaches obligations to each tier. So the first question is never "is it an agent," it is "what is the agent used for." This guide walks through how to answer that, what the obligations are, and the honest state of the timeline, which has already shifted once.

We build Pinchy, a self-hosted AI agent platform, and several of the Act's technical expectations overlap with what we build, so we are not neutral. This is a practical orientation, not legal advice; for a specific deployment, get advice that knows your situation.

Does it apply to you?

Two things decide it. First, reach: the Act applies extraterritorially, so if you provide or deploy an AI system whose output is used in the EU, you are in scope regardless of where your company sits. A self-hosted agent run by an EU company, or serving EU users, counts.

Second, risk tier. The Act sorts AI systems into four bands, and the band sets the obligations:

The practical takeaway: most internal business agents (summarizing support tickets, drafting copy, querying your own data) land in limited or minimal risk, where the burden is mainly being transparent that users are dealing with AI. An agent becomes high-risk because of its use case, such as making or materially informing decisions about employment, creditworthiness, or a regulated product, not because it is technically sophisticated. Classify the use, not the model.

The timeline, and why not to anchor on one date

The Act entered into force in 2024 and applies in phases. The phase that matters for most high-risk obligations was originally 2 August 2026, and this is where recent news matters: under the 2026 Omnibus package, obligations for use-based (Annex III) high-risk systems were postponed from 2 August 2026 to 2 December 2027, and product-related (Annex I) high-risk obligations from 2 August 2027 to 2 August 2028 (Gibson Dunn). General-purpose AI model obligations sit on their own schedule.

The useful lesson is not the new date, it is that the date moved. A regulation this large, operationalized through standards that are still being drafted, will keep adjusting. Building your agent program to a single deadline is fragile. Building it to the underlying requirements (which do not move) is not.

What high-risk actually requires

If an agent is high-risk, the obligation set is substantial, and most of it is recognizable as good engineering and governance:

• Record-keeping. Article 12 requires automatic logging of events over the system's lifetime, retained for at least six months, detailed enough to trace what happened. This is an audit trail requirement in all but name.
• Risk management. A standing process to identify and mitigate the system's risks, not a one-time assessment.
• Human oversight. A person able to understand, intervene in, and if needed stop the system. For an agent, that means a human can see what it is doing and pull it back.
• Bounded, documented operation. Technical documentation of what the system is, transparency to the people deploying it, and appropriate accuracy and robustness.

Read that list as an agent operator and it maps cleanly onto the four layers of agent governance: a record you can prove, access you can bound, a human who can intervene, and documentation of intent and ownership. The Act is, for agents, largely a legal restatement of controls a careful team would want anyway.

The penalties

The numbers are sized to be taken seriously, and they scale with the breach. A prohibited practice can draw up to 35 million euro or 7% of global annual turnover, whichever is higher. Non-compliance with high-risk or other obligations can reach up to 15 million euro or 3%. Supplying incorrect or misleading information to authorities can reach up to 7.5 million euro or 1%. The exposure is meaningful enough that "we will deal with it later" is itself a risk decision.

What to do now

Even with high-risk deadlines pushed toward late 2027, the move that ages well is to build the controls now, because they are the same controls that make an agent trustworthy independent of the law:

• Classify each agent's use case against the risk tiers, and write the classification down. Most will be low; the ones that are not are the ones to focus on.
• Turn on real logging. A tamper-evident, retainable audit trail satisfies the spirit of Article 12 and is useful long before any deadline.
• Bound access and keep a human in the loop for consequential actions, which is also just good agent design.
• Document what each agent is and who owns it, so the transparency and oversight obligations are a matter of record, not a scramble.

How Pinchy fits

This is the part about our own product, and the honest framing matters here more than anywhere: no platform makes you compliant. Compliance is a program (classification, risk management, documentation, human process) that is yours to run. What a platform can do is provide the technical building blocks so the engineering side is not the hard part. Pinchy gives you a tamper-evident audit trail that lines up with the record-keeping expectation, a default-deny permission model for bounded access, roles and ownership for accountability, and a self-hosted deployment that keeps the data inside your control. It does not classify your use cases or write your documentation, and it is not a certification. It makes the parts that are technical, technical, and leaves the parts that are judgment to you.