The EU AI Act for companies deploying AI agents: what applies, and when

A procurement questionnaire arrives with a new section: “Describe how your AI system complies with the EU AI Act.” The team that built the agent looks at the legal team, and the legal team looks back.

The short answer is that the AI Act asks different things of different systems, and of different companies using the same system. Two questions decide most of it: what risk class the use case is in, and whether you are the system’s provider or its deployer. This post walks through both, then the dates as they stand after the 2026 amendments, then five things worth doing now.

It is an engineer’s reading of the Regulation, written for the people who will ask their counsel to confirm it. We read every article cited here on 9 November 2026.

The four risk classes, in one paragraph each

Prohibited. A short list of practices is banned outright, such as manipulative techniques that cause significant harm and social scoring. These bans have applied since 2 February 2025. No business agent should come near them, but your classification should still say why it doesn’t.

High-risk. A system is high-risk in two ways. Either it is a safety component of a product covered by EU product law (Annex I: machinery, medical devices, toys and so on), or it is used in one of the areas listed in Annex III. The Annex III areas are biometrics, critical infrastructure, education, employment, access to essential private and public services, law enforcement, migration, and the administration of justice and democratic processes. For a company deploying agents, the entries in point 5 matter most: evaluating the creditworthiness of natural persons or establishing their credit score (fraud detection is excepted), and risk assessment and pricing for natural persons in life and health insurance. Our pages for consumer lending and insurance show where that line runs for agents in those sectors.

Transparency duties. Systems that interact with people must make sure “the natural persons concerned are informed that they are interacting with an AI system”, unless that is obvious (Article 50(1)). The information must come at the latest at the first interaction. Every customer-facing agent falls here, whatever its other risk class.

Minimal risk. Everything else. There are no specific AI Act duties, but GDPR, sector rules and your contracts still apply in full.

The filter that keeps many agents out of high-risk

Article 6(3) adds a filter: an Annex III system is not high-risk if it “does not pose a significant risk of harm to the health, safety or fundamental rights” of people. The article gives examples of when that is the case, such as a narrow procedural task, or improving the result of an activity a person has already completed.

Two limits make the filter less generous than it looks:

• A system that profiles natural persons is always high-risk.
• A provider who relies on the filter must document the assessment before placing the system on the market (Article 6(4)).

In practice this means the reasoning has to be written down for every agent, including the ones you conclude are not high-risk. Our free AI Act risk-class screener walks through these questions for one use case; it is a first screen, not legal advice. An agent that prepares a property insurance claim for a handler is not an Annex III use case at all. An agent that scores a loan applicant is, and the filter will rarely save it.

Provider or deployer: who carries which duty

The provider develops the system, or has it developed, and puts it on the market or into service under its own name. The deployer uses it under its own authority. When a specialist builds an agent for you to your specification, the contract should say who is which, because the duties differ.

For high-risk systems, most of the heavy requirements in Articles 9 to 15 fall on the provider: risk management, data governance, technical documentation, logging, transparency to deployers, human oversight by design, accuracy and robustness.

The deployer’s duties are in Article 26, and they are real:

• use the system according to its instructions;
• assign human oversight to people with “the necessary competence, training and authority”;
• monitor its operation and report serious risks and incidents;
• keep the logs under your control for at least six months, unless other law says otherwise (Article 26(6));
• inform the people affected when the system makes, or helps make, decisions about them (Article 26(11));
• inform workers’ representatives before using it at work.

Public bodies, private bodies providing public services, and deployers of the credit and life and health insurance systems in Annex III must also carry out a fundamental rights impact assessment before first use (Article 27). It can build on a GDPR data protection impact assessment you already have.

The dates, after the Digital Omnibus

The AI Act entered into force on 1 August 2024 and applies in stages. In November 2025 the Commission proposed the Digital Omnibus on AI, which changed several of those stages. According to the European Parliament’s research service, the co-legislators agreed the text on 7 May 2026 and Parliament’s plenary approved it on 16 June 2026.

The consolidated Article 113 now reads like this:

Two things follow for anyone deploying agents today. First, transparency is already a legal duty: a customer-facing agent that does not say it is an AI is out of line now, not in 2027. Second, the high-risk deadline moved, but the work did not get smaller. Building a credit or insurance-pricing agent to Articles 9 to 15 takes months, and an agent going live in 2026 will still be running in December 2027.

The Omnibus also softened the AI literacy duty in Article 4. It now asks providers and deployers to support literacy among their staff, with the Commission and Member States promoting it, rather than to ensure it. Ask your counsel to confirm the amending act’s Official Journal reference before you quote it in a contract.

Five things to do now

1. List your agents and classify each one. Include pilots and tools a vendor runs for you. For each, write down the use, the risk class and the reasoning. That list is the first thing an auditor will ask for.
2. Fix transparency today. Every agent that talks to customers or the public says it is an AI in its first message. This has applied since 2 August 2026.
3. Settle provider and deployer in your contracts. For every agent built for you, the contract should name the roles and say who keeps the logs, the documentation and the test results.
4. Put a person where it matters. Where an agent prepares decisions with legal or similar effects on people, design the approval step now. GDPR Article 22 already requires it for decisions based solely on automated processing, whatever the AI Act class. We wrote about designing that step.
5. Keep evidence as you go. Test results against an agreed bar, a record of every run, a record of who approved each release. These are hard to rebuild afterwards and cheap to keep from the start.

Where we stand

We build agents to the Article 9 to 15 standard where it helps, even when they are not high-risk, because it is a good engineering standard anyway. Our article-by-article map shows what our method and platform do for each requirement and what stays your job, and an AI Act readiness assessment writes the classification down for one of your use cases. None of it is a certificate. No certificate for being “AI Act compliant” exists, and anyone selling one is selling a badge, not evidence.

Questions

When do the AI Act's high-risk rules apply after the Digital Omnibus?

From 2 December 2027 for stand-alone systems listed in Annex III, and from 2 August 2028 for AI in products covered by Annex I, per the consolidated Article 113 we read on 9 November 2026. Counsel should confirm the dates against the Official Journal.

Is a customer service agent high-risk under the AI Act?

Usually not. It has the Article 50 transparency duty from 2 August 2026: people are told they are interacting with an AI. It becomes high-risk only if it is used in an Annex III area, such as evaluating the creditworthiness of people.

Can one company be both provider and deployer?

Yes. A company that has a system developed and puts it into service under its own name is its provider, and it is also the deployer when it uses it. The contract with the developer should say who carries which duty.

Sources

1. artificialintelligenceact.euartificialintelligenceact.eu/article/5/
2. artificialintelligenceact.euartificialintelligenceact.eu/article/6/
3. artificialintelligenceact.euartificialintelligenceact.eu/annex/3/
4. artificialintelligenceact.euartificialintelligenceact.eu/article/26/
5. artificialintelligenceact.euartificialintelligenceact.eu/article/27/
6. artificialintelligenceact.euartificialintelligenceact.eu/article/50/
7. artificialintelligenceact.euartificialintelligenceact.eu/article/113/
8. www.europarl.europa.euwww.europarl.europa.eu/RegData/etudes/BRIE/2026/782651/EPRS_BRI(2026)782651_EN.pdf

Read on or before 12 November 2026. Not legal advice: your counsel and data protection officer decide.

Written by

The Wise Agents team

Engineers who build and run AI agents for regulated companies. We read every article we cite on the day we state, and say what your counsel or DPO must confirm.

About us →