
This is general information, not legal advice.
The AI Act reaches beyond EU companies. It covers providers and deployers based outside the EU when their AI system's output is used in the Union (AI Act, Article 2). A UK retailer whose support bot answers customers in Germany is in scope.
The timetable changed this summer. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published on 24 July 2026 and entered into force on 27 July 2026 (Praxikon). It delayed the high-risk rules but left most transparency duties on their original date, 2 August 2026 (Jones Walker).
The EU AI Act dates that matter for AI agents
| Date | What applies | Who it affects most |
|---|---|---|
| 2 Feb 2025 | Banned practices, including emotion recognition in the workplace and in education; AI literacy duty | Anyone using AI with staff or candidates |
| 2 Aug 2025 | Rules for general-purpose AI models; penalty provisions | Model providers (OpenAI, Google, Anthropic and others) |
| 2 Aug 2026 | Article 50 transparency: tell people they are talking to AI; mark and label AI-generated content | Every business running a chatbot, voice agent or content tool for EU users |
| 2 Dec 2026 | Machine-readable marking for generative systems already on the market before 2 Aug 2026; new ban on systems that generate sexual abuse material or non-consensual intimate images | Providers of generative AI tools |
| 2 Dec 2027 | High-risk rules for Annex III uses, including recruitment, worker management and creditworthiness | Employers, staffing agencies, lenders |
| 2 Aug 2028 | High-risk rules for AI built into regulated products (Annex I) | Product manufacturers |
Sources: Future of Privacy Forum, Jones Walker, Praxikon.
The new dates are fixed in law and do not wait for technical standards (Praxikon).
Article 50: telling people they are talking to AI
Article 50(1) requires AI systems that interact with people to be designed so that people know they are dealing with AI. The only general exception is where that is already obvious to a reasonably well-informed, observant person, and commentators warn against relying on it (artificialintelligenceact.eu).
The European Commission published its final guidelines on Article 50 on 20 July 2026 (European Commission). As summarised by Stephenson Harwood, the duty applies when four things are true: it is an AI system, it is meant to interact with people, the interaction is direct (real time or near real time), and the other side is a person rather than another machine. Rule-based automation is outside it.
What good disclosure looks like:
- When: at the latest at the first interaction, not after the customer has shared details.
- How: clear, distinguishable and accessible. A notice at the start of the chat, a persistent "AI assistant" badge, or a spoken line at the start of a call all work.
- What does not work: a line in the terms and conditions, faint footer text or a label that flashes past (artificialintelligenceact.eu).
- AI agents that act for you: an agent acting on its own should say on whose behalf it acts, and repeat this at key steps such as taking a payment authorisation or reporting back (Stephenson Harwood).
For a voice agent, a greeting such as "Hi, you're speaking with the practice's AI assistant. I can book or move appointments, or put you through to the team" covers it in one sentence.
Article 50(1) is a duty for the provider, the company that builds the system or puts it on the market, but provider and deployer duties can apply together, so record which role you hold for each system (Jones Walker). If you commission a custom agent and run it under your own brand, ask your lawyer whether you count as its provider. Either way, build disclosure in.
AI-generated content: marking and labelling
If you use generative AI for content, two more duties apply. Providers of tools that generate text, images, audio or video must mark the output as AI-generated in a machine-readable way; tools already on the market before 2 August 2026 have until 2 December 2026 (Jones Walker). Businesses must label deepfakes and AI-generated text published on matters of public interest, unless a person has reviewed it and holds editorial responsibility (artificialintelligenceact.eu).
The Commission's voluntary Code of Practice on marking and labelling, published on 10 June 2026, sets out practical steps (European Commission).
Hiring AI: what applies now and what waits until December 2027
Already in force:
- No emotion reading. Emotion recognition in the workplace has been banned since February 2025, with narrow exceptions (artificialintelligenceact.eu). Treat AI interviews as covered: a voice or video interviewer should not score candidates on inferred confidence, enthusiasm or stress.
- Disclosure. A voice interviewer or screening chatbot is an AI system talking to a person, so Article 50 applies: candidates must know they are speaking with AI.
- AI literacy. Since the Omnibus, employers must take measures that support AI literacy among staff who use these tools (Future of Privacy Forum).
From 2 December 2027: AI used for recruitment, selection, targeted job ads, candidate evaluation, performance monitoring and some decisions on contract terms or termination is high-risk. Employers using it as deployers will need to (artificialintelligenceact.eu):
- assign trained people who can spot and correct errors, including discriminatory patterns;
- keep the input data they control relevant and representative;
- monitor how the system performs;
- keep the logs it generates for at least six months;
- inform workers' representatives and affected workers before deployment;
- explain the main factors behind a decision when a candidate asks.
Which AI agent jobs fall where
| Agent job | Applies now | Extra from 2 Dec 2027 |
|---|---|---|
| AI receptionist or voice agent answering calls | Article 50 disclosure at the start of each call | None, unless it makes high-risk decisions |
| Website or WhatsApp support chatbot | Article 50 disclosure in the first message | None, unless it makes high-risk decisions |
| Lead qualification or appointment setting | Article 50 disclosure; say on whose behalf the agent acts | None in most cases |
| Payment reminder or collections agent | Article 50 disclosure | High-risk only if it is used to assess creditworthiness |
| AI interview screening | Article 50 disclosure; no emotion recognition | High-risk duties: oversight, logs, worker information, explanations |
| AI-written blog posts, ads or images | Deepfake labels; public-interest text labels unless human-edited | None |
A starting map only: classification depends on each system's purpose, so check it with your lawyer.
A seven-step checklist for businesses using AI agents with EU users
- List every AI system that talks to customers, candidates or staff in the EU, including vendor tools.
- For each one, record whether you are the provider, the deployer or both, and why each duty does or does not apply (Jones Walker).
- Add a clear AI disclosure at the first interaction: the first chat message, the call greeting, the interview invitation.
- For agents that act on your behalf, state whose behalf, and repeat it before payments, bookings or other commitments.
- Remove any feature that infers emotions from staff or candidates.
- Label deepfakes and unreviewed AI-written public-interest content; keep a named editor for content you publish.
- If you use AI in hiring or credit decisions, start building oversight, logging and worker information now, ready for 2 December 2027.
Five questions to ask your AI agent vendor
- How does the agent tell users it is an AI, and can we change the wording and language?
- Does the disclosure work in every channel we use: web chat, WhatsApp, phone?
- Does any feature analyse voice or video to infer emotions? Can it be switched off completely?
- What does the agent log, where are the logs stored, and can we keep them for at least six months?
- How does a person take over a conversation, and does the agent pass on the full history?
How Prilient can help
Prilient builds custom AI agents and runs Prilient AI Agents: Receptionist, Customer Support, WhatsApp Lead Generation, Appointment Setter, Debt Collection and HR Interview Screening. For AI voice agents and chatbots serving EU users, we plan the disclosure wording, the human handoff and the logging with you before the agent goes live. Through AI consulting, we map the AI tools you already use against these dates.
Book a free 30-minute call. Tell us which agents you run and where your users are, and we will show you what to change first.
Questions about the EU AI Act and AI agents
Does the EU AI Act apply to my business if we are not based in the EU?
It can. The Act covers providers and deployers outside the EU when the output of their AI system is used in the EU.
Do AI chatbots have to say they are AI?
Yes, since 2 August 2026, unless it is already obvious to a reasonably well-informed person. The disclosure must be clear and given at the latest at the first interaction. A line in the terms and conditions is not enough.
How should an AI voice agent disclose that it is AI?
With a spoken line at the start of the call, before the caller shares details: "You're speaking with our AI assistant. I can help with bookings, or connect you to the team."
When do the EU AI Act rules for AI in recruitment start?
The high-risk duties for recruitment and other employment uses apply from 2 December 2027, after the Digital Omnibus moved them from 2 August 2026. The ban on emotion recognition at work and the AI disclosure duty already apply.
Can an AI interview tool analyse a candidate's tone or facial expressions?
Not to infer emotions. Emotion recognition in the workplace has been banned since February 2025, with narrow exceptions, and employers should treat AI interviews as covered.
What are the fines for not disclosing that a chatbot is AI?
Article 50 breaches carry fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs and startups the lower figure applies, and fines must be proportionate.
Leave A Reply
Your email address will not be published. Required fields are marked *