For AI agents in customer service, the EU AI Act’s main rule is transparency: since August 2, 2026, people must be told they are interacting with an AI, from the first interaction, unless that is obvious.
Most support chatbots and voice bots are not high-risk. AI that monitors and evaluates your human agents can be, and inferring employees’ emotions from biometric data is banned.
In short
- Article 50(1) of the EU AI Act requires that customers are told they are talking to an AI, at the latest at the first interaction, and has applied since August 2, 2026.
- The Digital Omnibus on AI, in force since July 27, 2026, moved the high-risk dates to December 2, 2027 and August 2, 2028, but did not delay the chatbot rule.
- A customer-facing support AI is usually in the transparency tier, while AI used to monitor and evaluate employees falls under a high-risk category unless an exception applies.
- AI that scores your human agents raises two questions: the workplace emotion recognition ban, and the high-risk rules for monitoring workers.
- QA can show compliance on every conversation: score the AI disclosure, the handoff to a person and what the AI agent promised.
This article explains the law as it stands on November 4, 2026. It is not legal advice. For the QA side of AI agents, the page on AI agent QA shows how Kaizo scores AI agent conversations on the same scorecard as your team’s.
What does the EU AI Act require of AI agents in customer service?
The EU AI Act requires that an AI agent talking to customers is designed so they know it is an AI, unless that is obvious from the context. This is Article 50(1), and it is a duty of the provider.
Article 50(1) of Regulation (EU) 2024/1689 says providers must ensure that AI systems intended to interact directly with people are designed so that those people are informed they are interacting with an AI system, unless this is obvious to a reasonably well-informed, observant and circumspect person.
Article 50(5) adds that the information must be clear and distinguishable, given at the latest at the time of the first interaction, and accessible.
The European Commission’s FAQ on Article 50, updated July 24, 2026, adds detail that matters for support teams:
| Point | What the Commission says |
|---|---|
| Which systems | Four conditions together: an AI system, a genuine two-way exchange, the AI communicates directly with no human in between, and the other side is a person. Chatbots, AI agents and avatars are given as examples. |
| What falls outside | Systems that work only in the background or machine to machine. |
| The “obvious” exception | Read restrictively, because it deprives people of transparency. |
| Timing | People must be notified from the start of the first interaction. |
| Form | Providers choose the technique, but a notice only in the terms and conditions is not enough on its own. |
| Guidance | Commission guidelines on Article 50 were adopted on July 20, 2026 (C(2026) 5054). |
| Enforcement | Mainly national market surveillance authorities. |
Who is the provider? Usually the company that builds the AI agent, which may be your helpdesk vendor or your own team if you build it yourself.
If you deploy a vendor’s AI agent, you still own what your customers see: check that the disclosure is switched on in every channel and every language you run, because a customer does not care whose duty it was.
Key dates of the EU AI Act for support teams
The AI Act entered into force on August 1, 2024 and applies in stages; the Digital Omnibus on AI moved the high-risk dates and left the chatbot rule on August 2, 2026.
| Date | What applies | Relevance for support |
|---|---|---|
| August 1, 2024 | The AI Act enters into force | None directly |
| February 2, 2025 | Prohibited practices (Article 5) and AI literacy (Article 4) | The ban on inferring employees’ emotions from biometric data |
| August 2, 2025 | Rules for general-purpose AI models, governance, penalties | Mostly for model providers |
| July 27, 2026 | The Digital Omnibus on AI, Regulation (EU) 2026/1744, enters into force | Moves the high-risk dates below |
| August 2, 2026 | General date of application, including Article 50 | AI disclosure for chatbots and voice bots |
| December 2, 2026 | Machine-readable marking (Article 50(2)) for generative AI systems placed on the market before August 2, 2026 | Mostly for vendors of generative AI |
| December 2, 2027 | High-risk rules for Annex III systems, including employment | AI used to monitor and evaluate workers |
| August 2, 2028 | High-risk rules for Annex I systems (AI in regulated products) | Rarely relevant to support |
Sources: the dates come from Articles 111 and 113 of Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, checked against the European Commission’s AI Act page, updated October 19, 2026. The Omnibus was published in the Official Journal on July 24, 2026.
It also rewrote Article 4: providers and deployers must now take measures to support the AI literacy of their staff, and the text says this does not require them to guarantee any specific level.
Which risk class does a customer service AI fall into?
A customer-facing support AI usually falls into the transparency tier, with the disclosure duty and nothing more. Classification depends on what the system is intended for, not on the technology.
The AI Act sorts systems into four tiers. Here is how typical support uses map onto them, as a starting point for your own assessment:
| Tier | What it means | Support examples |
|---|---|---|
| Prohibited (Article 5) | Banned since February 2, 2025 | Inferring an agent’s emotions from their voice or face at work, outside medical or safety reasons |
| High-risk (Annex III) | Strict obligations, from December 2, 2027 for Annex III | AI used to make decisions on agents’ work terms, promotion or termination, or to monitor and evaluate their performance and behavior (point 4(b)); credit scoring of customers |
| Transparency (Article 50) | Disclosure duties, since August 2, 2026 | A chatbot, voice bot or AI agent that talks to customers |
| Minimal | No specific rules | Spam filtering, ticket routing by topic, spelling suggestions |
Two details change the picture.
First, Article 6(3) says an Annex III system is not high-risk if it poses no significant risk of harm, for example because it performs a narrow procedural task or a preparatory task to a human assessment, but a system that profiles people is always high-risk.
Second, the tiers stack: a high-risk system that also talks to people must meet Article 50 as well.
Most support AI agents answer questions and take simple actions. If yours also decides on refunds above a limit, credit or access to essential services, review that use case against Annex III point 5 with your counsel.
What about AI that scores your human agents?
AI that scores your agents raises two questions: does it infer emotions from biometric data, which is banned at work, and does it evaluate performance, a high-risk use unless an exception applies?
The emotion recognition ban. Article 5(1)(f) prohibits using AI systems to infer the emotions of a person in the workplace, except for medical or safety reasons, and has applied since February 2, 2025.
The Act defines an emotion recognition system in Article 3(39) as one that infers emotions or intentions on the basis of biometric data, such as a voice or a face. Recital 18 adds that merely detecting a raised voice is not emotion recognition, unless it is used to infer emotions.
The Commission’s guidelines on prohibited AI practices, published on February 4, 2025 and not binding, use a call center as an example.
As summarized by the law firm Lewis Silkin, tracking employees’ emotions with voice recognition to monitor how call center staff deal with customers is given as prohibited, while inferring emotions from written text is not based on biometric data and falls outside the ban.
What this means in practice:
| Use | Status under the AI Act |
|---|---|
| Inferring an agent’s anger or stress from their voice on calls | Prohibited at work, unless for medical or safety reasons |
| Sentiment of the customer, read from the text of a chat or email | Not emotion recognition (no biometric data) |
| Emotions of the customer, inferred from their voice | Not covered by the workplace ban, but an emotion recognition system: listed as high-risk in Annex III point 1(c), and Article 50(3) requires informing the people exposed to it |
| Checking whether the agent followed the procedure or answered correctly | Not emotion recognition |
The high-risk question. Annex III point 4(b) lists AI systems intended to monitor and evaluate the performance and behavior of people in work relationships. A QA tool that scores agents can fit that description, unless an Article 6(3) exception applies, for example where the AI prepares an assessment a person makes.
These obligations apply from December 2, 2027. Deployers that are employers must then inform workers’ representatives and the affected workers before putting a high-risk AI system into use at work (Article 26(7)).
How a given QA tool is classified depends on its intended purpose and how you use it. Ask your vendor for its own assessment, and involve your works council or employee representatives early: in several EU countries, monitoring tools need their agreement anyway.
For how scoring with human review works in practice, see our guide on how to validate AI QA scoring.
How can QA show that your AI agent complies?
QA shows compliance by scoring every AI agent conversation for what the law and your policy require, such as the AI disclosure and a route to a person, with each score linked to evidence.
Article 50 sets a rule for the design of the system. Whether it works in production, in every channel, language and flow, is something you only know by checking the conversations. Add these criteria to your AI agent scorecard:
| Criterion | What it checks | Type |
|---|---|---|
| AI disclosure | The customer was told at the start that they were dealing with an AI, in the conversation’s language | Auto-fail |
| Disclosure on handback | If a person hands the conversation back to the bot, the customer knows | Weighted |
| Route to a person | When the customer asked for a human, they got one or a clear way to reach one | Auto-fail |
| Handoff context | The person who took over had the conversation history (how to score the handoff) | Weighted |
| No invented commitments | No refund, fee waiver or policy the company does not offer | Auto-fail |
| Grounded in your content | Facts come from your own knowledge base | Weighted |
The route to a person is not an AI Act duty for every chatbot, but it is where customer complaints about AI agents start, and our guide on silent failures in AI agents shows what goes wrong when the bot never escalates. For the full method, see how to QA AI agents.
See it in a live demo. Kaizo scores 100% of your AI agent conversations against your own scorecard and links every score to the evidence in the conversation. Let’s talk.
What are the fines under the EU AI Act?
Breaching the Article 50 transparency duties can lead to fines of up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher, under Article 99(4) of the AI Act. Prohibited practices carry the highest ceiling, up to EUR 35 million or 7% under Article 99(3).
The same ceiling applies to the obligations of deployers of high-risk systems under Article 26.
For SMEs, including start-ups, Article 99(6) caps each fine at whichever of the two amounts is lower. National authorities set the actual fine, and the Commission FAQ notes that proportionality is considered for SMEs and small mid-cap companies.
How to prepare your support team for the EU AI Act
Prepare in five steps: list your AI systems, classify each by use, check the AI disclosure in every channel, settle how AI scoring of agents works, and score every AI agent conversation for the disclosure.
- List every AI system in support. Chatbots, voice bots, AI agents, agent assist, QA scoring, sentiment, routing. Note who provides each one and what it is used for.
- Classify each by intended use. Use the risk table above as a first pass, then confirm with counsel. Mark anything that evaluates employees or decides on customers’ access to services.
- Check the disclosure in every channel and language. Open a conversation with each bot as a customer would, in chat, email, voice and every language, and check the first message.
- Settle how AI scores your agents. Confirm that no tool infers agents’ emotions from voice or video, decide where a person reviews AI scores, and inform employee representatives.
- Score every AI agent conversation for the disclosure. A one-time check proves the design; scoring all conversations shows it held after the next flow change. That is what QA for AI agents in Kaizo is built to do.
Which law applies? Regulation (EU) 2024/1689 (the AI Act), as amended by Regulation (EU) 2026/1744 (the Digital Omnibus on AI), checked on EUR-Lex and the European Commission’s AI Act pages on November 4, 2026.
Commission guidelines and codes of practice are not binding and may still change, and national authorities may publish their own guidance, so check before you act on a specific case.
Frequently asked questions
Does a customer service chatbot have to say it is an AI?
Yes, in the EU, since August 2, 2026. Article 50(1) of the AI Act requires that people are told they are interacting with an AI system, at the latest at the first interaction, unless that is obvious. The Commission reads the “obvious” exception restrictively, so a clear notice at the start of the conversation is the safe choice.
Is a customer service chatbot a high-risk AI system?
Usually not. A chatbot that answers support questions falls under the transparency rules of Article 50. It can become high-risk if it is used for a purpose listed in Annex III, for example evaluating creditworthiness or deciding on access to essential services. Classification depends on the intended purpose, so check each use case.
Did the Digital Omnibus delay the AI Act’s chatbot rules?
No. The Digital Omnibus on AI, Regulation (EU) 2026/1744, moved the high-risk dates to December 2, 2027 and August 2, 2028. Article 50 kept its date of August 2, 2026. The Omnibus only gave providers of generative AI systems already on the market until December 2, 2026 for the machine-readable marking duty in Article 50(2).
Does the EU AI Act apply to companies outside the EU?
Yes, in many cases. Under Article 2, it applies to providers placing AI systems on the EU market, whatever their location, and to providers and deployers outside the EU when the output of the system is used in the EU. A US company whose AI agent serves customers in the EU should assume the disclosure rule applies.
Is sentiment analysis of support conversations allowed under the AI Act?
Sentiment read from chat and email text is not emotion recognition: it uses no biometric data (see sentiment scoring in QA). Inferring an employee’s emotions from voice or face at work is prohibited, except for medical or safety reasons. Inferring a customer’s emotions from their voice is high-risk under Annex III, and they must be informed (Article 50(3)).
