The current deadlines, when an agent is high-risk and what your organisation records.
Yes. But agents for customer questions, email or invoices are usually not high-risk systems. Then it is mainly about transparency: people must know they are talking to AI. That duty has applied since 2 August 2026. The rules for high-risk systems such as screening job applicants apply from 2 December 2027.
The Digital Omnibus (Regulation 2026/1744) entered into force on 27 July 2026 and moved some of the dates.
| Date | What applies | What it means for an agent |
|---|---|---|
| 2 February 2025 | Prohibited AI practices and AI literacy | Manipulation and certain forms of profiling are prohibited. Staff working with the agent need to understand it sufficiently. The Omnibus softened the literacy duty. |
| 2 August 2026 | Transparency (Article 50) | Anyone talking to an agent must know it is AI, unless that is obvious from the context. |
| 2 December 2026 | Marking of AI-generated content | Transition period for systems on the market before 2 August 2026. The duty lies with the provider of the system. |
| 2 December 2027 | High-risk (Annex III) | For example agents for recruitment, credit scoring or access to services. |
| 2 August 2028 | High-risk in products (Annex I) | AI as part of regulated products, such as medical devices. |
Not the technology but the use determines the risk class. The same agent can therefore move to another class when its task changes.
| Use | Usually | Why |
|---|---|---|
| Customer questions, mailbox and appointments | Transparency duty | The agent talks to people, but makes no decision that determines their rights or access. |
| Invoices, quotes and reports | No specific AI Act duty, or transparency only | Administrative work, as long as the agent does not assess people. |
| Selecting or assessing job applicants | High-risk | Recruitment and staff management are listed in Annex III. |
| Assessing creditworthiness | High-risk | Listed in Annex III. Exception: detecting financial fraud. |
| Risk and pricing in life and health insurance | High-risk | Listed in Annex III. |
| Access to public services or emergency healthcare triage | High-risk | Listed in Annex III. |
This is an indication. The final classification depends on the actual use and we record it per agent.
For a limited-risk agent, this is a practical minimum:
Record what the agent does, which data it uses and which risk class it falls into
Tell users at the start of the conversation that they are talking to AI
Decide who is ultimately responsible and when the agent hands over to a person
Log what the agent does, so you can trace what happened later
Make sure staff know what the agent can and cannot do
Carry out a DPIA if the agent processes personal data with a high privacy risk; the GDPR applies alongside the AI Act
Every agent comes with an AI Act file and a DPIA by default. The agent says it is AI, every action is logged, hosting and data are in the EU, and when in doubt the agent stops and a person steps in.
Does your legal team remain ultimately responsible? Yes. The file gives them the information to assess the agent.
This page is a practical summary and not legal advice. Have a lawyer assess how it applies to your situation.
Question not listed? Ask it in a process call
Yes. Since 2 August 2026, Article 50 of the AI Act requires this, unless it is obvious to the user from the context. We build the notice in by default.
Usually not. That can change if the agent decides whether someone gets access to an essential service. That is why we record the task per agent.
Yes. Under the Digital Omnibus they apply from 2 December 2027 for Annex III and from 2 August 2028 for Annex I. The transparency duty for AI interaction was not postponed.
Up to 35 million euros or 7 percent of worldwide annual turnover for prohibited practices, and up to 15 million euros or 3 percent for most other obligations. For SMEs the lower of the two amounts applies.
No. If the agent processes personal data, both apply. The DPIA belongs to the GDPR, the AI Act file to the AI Act.
In a 30-minute process call we look at the task of the agent and what you need to record for it.