The EU AI Act and AI agents: what changes for you

The AI Act, the European regulation on artificial intelligence, entered into force on 1 August 2024 and applies in stages. In July 2026, an amending regulation, the Digital Omnibus on AI, shifted part of the timeline. Many companies deploying agents are therefore asking what already applies, what is coming, and what it changes for them. Here is our reading, as the maker of an agent platform.

This article is not legal advice. The dates and articles quoted were checked on 30 September 2026 against the official texts, listed at the end; check them again when you read this, and have your situation reviewed by your lawyer or DPO.

The regulation in brief

The AI Act classifies uses of AI by level of risk, and sets proportionate obligations. Some practices are prohibited, such as social scoring or emotion recognition in the workplace, except for medical or safety reasons. High-risk systems, listed by the regulation, are subject to strict requirements. Other systems are subject to transparency obligations. The rest, which covers most business uses, has no specific obligation beyond the general rules.

The key point for agents: it is the use that determines the risk, not the technology. The same agent summarising meeting notes is not affected in the same way as if it were screening job applications.

The application timeline, after the Omnibus

Regulation (EU) 2026/1744 of 8 July 2026, known as the Digital Omnibus on AI, in force since 27 July 2026, postponed the obligations for high-risk systems. The timeline published by the European Commission is now as follows:

The postponement buys time, but exempts no one: prohibited practices and AI literacy already apply, and high-risk uses need preparing now, because they require reworking processes, not just a tool.

Provider or deployer?

The regulation distinguishes two main roles. The provider develops an AI system, or has it developed, and places it on the market under its own name. The deployer uses an AI system under its own authority, in a professional context. A company that uses an agent platform for its teams is, as a rule, a deployer.

Be careful, though: Article 25 provides that a deployer becomes the provider of a high-risk system if it puts its name or trademark on it, substantially modifies it, or changes its intended purpose so that it becomes high-risk. Configuring an agent to screen applications, when the tool was not intended for that, may be enough to change roles.

When does an agent become “high-risk”?

Annex III lists the areas where an AI system is presumed high-risk. For a company, the most common are the following:

• Employment and workers’ management: recruitment, screening applications, evaluating candidates, promotion or termination decisions, allocating tasks based on behaviour, monitoring performance.
• Access to essential services: assessing people’s creditworthiness or credit score, risk assessment and pricing in life and health insurance.
• Education and training: admission, assessing learners, monitoring exams.

The regulation also provides exceptions: a system listed in Annex III is not considered high-risk if it does not pose a significant risk, for instance because it performs a narrow procedural task or a task preparatory to a human assessment. These exceptions do not apply when the system profiles people. An agent that prepares a file for a recruiter is therefore not in the same situation as an agent that ranks candidates.

The deployer’s obligations

For all uses, two obligations already apply. Prohibited practices are banned. And, under Article 4, providers and deployers take measures to develop the AI literacy of their staff and of the people who use the systems on their behalf. Since 2 August 2026, some transparency obligations are added, for instance disclosing the deepfakes you publish.

For a high-risk use, Article 26 adds specific obligations:

• Use the system in accordance with its instructions for use.
• Assign human oversight to people who have the necessary competence, training and authority.
• Ensure the input data is relevant, to the extent you control it.
• Monitor how the system operates, suspend its use in case of risk, and inform the provider and the authority of any serious incident.
• Keep the automatically generated logs under your control for at least six months, unless otherwise provided.
• Inform workers’ representatives and the affected workers before putting the system into service at the workplace.
• Inform people who are subject to a decision made or assisted by the system.

Public bodies, private companies providing public services, and deployers that assess creditworthiness or price life and health insurance must also carry out a fundamental rights impact assessment before putting the system into service (Article 27).

Penalties are set by Article 99: up to €35 million or 7% of worldwide turnover for prohibited practices, up to €15 million or 3% for most other obligations, including those of deployers. For SMEs, the lower of the two amounts applies.

The checklist

1. List your AI uses, agents included, with what each does and for whom.
2. Check that none falls under a prohibited practice.
3. Identify those that touch an Annex III area, and document why they are, or are not, high-risk.
4. Check your role: deployer, or provider if you changed a tool’s intended purpose.
5. Set up AI training for the teams that use agents.
6. For high-risk uses: name the people in charge of oversight, organise log retention, plan how employees and affected people will be informed.
7. Ask your vendors for the instructions for use and documentation that let you meet your obligations.

What the platform must let you do

The deployer’s obligations remain yours; no tool fulfils them for you. But an agent platform can make them practical, or impossible. For human oversight, you need to be able to require approval before the actions that matter, and that approval has to actually block. For logs, every execution has to be traced: the request, the sources read, the steps, the approvals. For input data, you need to know what the agent reads and be able to exclude what is not relevant.

In Scalt, you define the approval steps, an action subject to approval stops and waits for your consent, every execution is traced in Cases, and each agent only reads what it is allowed to read. It is a working basis for your compliance, not turnkey compliance.

Frequently asked questions