If you run operations or compliance in a company that uses an AI agent, the EU AI Act (Regulation (EU) 2024/1689) is probably on your list, and probably with a question mark. The text is long, the dates have changed since it was adopted, and most summaries are written for the companies that build AI systems, not for those that use them.
This article is for the second group. It explains which parts of the Act are likely to touch you, in what order to check them and what to prepare. It is an orientation for a non-lawyer, based on the official texts linked at the end. It is not legal advice, and where your situation is unusual you should ask a lawyer who knows your sector.
Where the dates stand
The Act entered into force on 1 August 2024 and applies in stages. According to the European Commission, the picture today is this:
- 2 February 2025: the general provisions, the AI literacy duty and the prohibited practices apply.
- 2 August 2025: the rules on general-purpose AI models and the governance structure apply.
- 2 August 2026: most of the remaining rules apply, including the transparency rules in Article 50, and enforcement starts.
- 2 December 2027: the rules for high-risk AI in the sensitive areas listed in Annex III, such as employment, education, biometrics or critical infrastructure.
- 2 August 2028: the rules for high-risk AI built into regulated products such as lifts or toys.
The last two dates are later than the original text, because the so-called AI Omnibus amendment entered into force on 27 July 2026. If you read an older article that gives August 2026 for high-risk systems, check it against the Commission's current timeline. Dates in this area have moved once and may be adjusted again by guidelines or further amendments, so keep one person responsible for re-checking them.
Step one: which role do you have?
The Act gives different duties to different roles. The two that matter here are the provider, who develops an AI system or has it developed and places it on the market under its own name, and the deployer, who uses an AI system under its own authority in a professional context.
If you buy an agent or an agent platform and use it in your business, you are in most cases a deployer. Your vendor is the provider. This matters because most of the heavy obligations, such as the risk management system, technical documentation and conformity assessment, sit with the provider.
There is an important exception. Article 25 says that a deployer is treated as a provider of a high-risk system if it puts its own name or trademark on such a system, makes a substantial modification to it, or changes the intended purpose of a system that was not high risk so that it becomes high risk. A company that takes a general-purpose agent and sets it to rank applicants has not simply "used" a tool. It may have taken on the provider's role.
So the first practical question is not "is our agent high risk?" but "who is the provider, and have we changed what the system was meant for?" This is also a contract question; see what to look for in AI agent vendor contracts.
Step two: classify each use, not the agent
The Act does not classify technology; it classifies uses. One agent platform can serve several processes with very different risk. We suggest a simple inventory with one row per process:
- What does the agent do, and with whose data?
- Who is affected by its output - customers, candidates, employees, the public?
- Does the output influence a decision about a person?
- Which risk label applies, and who decided it?
The labels you will meet are these.
Prohibited practices (Article 5). A short list of uses that are banned, such as certain forms of manipulation or social scoring. A normal business agent is unlikely to be anywhere near it, but read the list once and record that you did. Fines for this category are the highest: up to EUR 35 million or 7% of worldwide annual turnover, whichever is higher.
High-risk uses (Article 6 and Annex III). The Commission names areas such as biometrics, critical infrastructure, education, employment, migration and border control. In practice, an operations lead should look hard at any agent that screens or ranks job applicants, allocates tasks or monitors performance of staff, or assesses people for access to essential services. Those are the situations where an everyday productivity tool crosses into a regulated category.
Transparency-only uses (Article 50). This is where most customer-facing agents live. See the next section.
Everything else. Many internal uses - drafting, summarising, searching documents, preparing reports - fall outside the specific rules. They still fall under the AI literacy duty and under your other laws, above all data protection and contracts.
If you cannot decide between two labels, write down why and treat the stricter one as the working assumption until someone with the right expertise says otherwise.
Transparency: what you tell people
Article 50 has been in application since 2 August 2026. For a company with a customer-facing agent, three of its rules matter.
- Telling people they are talking to an AI. The provider must design the system so that people are informed they are interacting with AI, unless that is obvious to a reasonably well-informed person. In practice, you should still make sure that your own chat window, voice line or email signature does not hide it. A short notice and a clear way to reach a person is a sensible default; our article on what to say to customers and when goes through wording.
- Emotion recognition and biometric categorisation. Deployers of such systems must inform the people exposed to them. If an agent product offers sentiment or voice analysis of customers or staff, find out whether it falls into this category before you switch it on.
- Synthetic content. Deployers who generate deepfake image, audio or video must disclose that it is artificial. Text generated by AI and published to inform the public on matters of public interest must be disclosed too, unless it has been through human review or editorial control and someone holds editorial responsibility. Marketing teams that use agents to produce public content should know which side of that line they are on.
Transparency duties for providers, such as marking synthetic output in a machine-readable way, are mostly the vendor's job. Ask the vendor how it meets them rather than assuming.
Human oversight and AI literacy
Two duties apply broadly and are easy to prepare.
AI literacy (Article 4). Providers and deployers must take measures to support the AI literacy of their staff and others who operate AI systems on their behalf, taking into account their knowledge, experience and the context of use. As amended, the text states that this does not require guaranteeing any specific level of literacy for any individual. In practice that means documented, proportionate training for the people who work with the agent: what it can and cannot do, when to stop it, what to report.
Human oversight (Article 26 for high-risk systems). Deployers must assign human oversight to natural persons who have the necessary competence, training, authority and support. A named person with no time and no power to stop the agent does not meet the spirit of this. The same idea is good practice for every agent; see our articles on autonomy levels and on human approval and audit trail.
Records and logs
For high-risk systems, Article 26 requires the deployer to keep the logs automatically generated by the system, to the extent they are under its control, for a period appropriate to the intended purpose and at least six months, unless other Union or national law says otherwise, especially data protection law. The practical consequences:
- Find out whether your vendor produces such logs at all and whether you can export them.
- Decide who owns them, where they are stored and who can read them.
- Reconcile the retention period with your data protection obligations; a longer log is not automatically better when it contains personal data.
Even outside high-risk uses, a log is how you answer the question "what did the agent do and why?" after a complaint. Our guide to the activity log lists what is worth recording.
Other deployer duties for high-risk uses
If one of your uses is high risk, Article 26 adds further tasks. Deployers must:
- use the system according to the provider's instructions for use;
- ensure that input data under their control is relevant and sufficiently representative for the intended purpose;
- monitor operation and, if they have reason to think the system presents a risk, inform the provider and the relevant authority and suspend use; serious incidents must be reported to the provider first and then to the authorities;
- inform workers' representatives and affected workers before using a high-risk system at the workplace;
- inform people who are subject to decisions, or assisted decisions, made with an Annex III high-risk system.
Some deployers, such as public bodies or private entities providing public services, have further duties, for example a fundamental rights impact assessment under Article 27 and registration. Whether that applies to you depends on your sector, which is one of the points to bring to a lawyer.
Notice how close this is to ordinary good operations: instructions followed, monitored, incidents reported, people told. If you already have incident response and kill switches and a monthly review, much of the groundwork exists.
Penalties
Member States set the rules on penalties within limits given in Article 99. For deployer obligations under Article 26 and transparency obligations under Article 50, the ceiling is EUR 15 million or 3% of total worldwide annual turnover, whichever is higher, for an undertaking. For small and medium-sized enterprises the lower of the two amounts applies. The ceiling is not what a typical fine will be, but it shows that regulators regard these duties as real.
A checklist to start this month
- Inventory every process where an AI agent or AI feature is used, including features inside tools you did not think of as AI.
- Label each use as prohibited, high risk, transparency-relevant or other, with a short reason and a name.
- Confirm your role for each: deployer, or possibly provider because you changed the purpose or modified the system.
- Ask your vendors for the instructions for use, the log export, their position on Article 50 and their own plan for the Annex III dates.
- Set up training that you can show: who was trained, on what, and when.
- Name the oversight person for every high-risk or customer-facing use, with the authority to pause the agent.
- Check the notices customers and staff see, and the contract clauses on liability and exit.
- Put a review date in the calendar, because the timeline and the Commission's guidance are still developing.
If your inventory turns up a possible high-risk use, this is the point to involve legal counsel, well ahead of December 2027. Everything else on the list is a few weeks of work for one organised person, and it also makes your AI governance better, whatever the regulators finally decide.
Sources
- European Commission: AI Act - application timeline and the AI Omnibus
- AI Act Service Desk: timeline for the implementation of the EU AI Act
- Regulation (EU) 2024/1689 on EUR-Lex
- AI Act Service Desk: Article 26, obligations of deployers of high-risk AI systems
- AI Act Service Desk: Article 50, transparency obligations