Chatbots and Voicebots: the Duty to Tell Customers They Are Talking to AI

What Article 50 says, paragraph by paragraph#

Article 50 of Regulation (EU) 2024/1689 groups together several transparency obligations. They do not all target the same operator, and they do not all affect a contact centre in the same way.

Article 50 has applied since 2 August 2026. The "Digital Omnibus on AI" (Regulation (EU) 2026/1744) did not postpone it. Only the marking under Article 50(2) benefits from extra time, until 2 December 2026, for generative systems placed on the market before 2 August 2026. The regulation's wider context for the sector is set out in our guide to the AI Act and Moroccan BPOs.

Who must inform: the software vendor, the client or the centre?#

Legally, Article 50(1) falls on the provider, meaning whoever develops the system and places it on the market under its own name. In practice, it is often the call centre that configures the welcome message, the voicebot script or the chat banner. Three situations come up repeatedly:

• The centre has developed its own voicebot and offers it to its clients: it is the provider and carries the obligation directly.
• The centre uses a software vendor's tool: the vendor must make the information possible, but the centre must make sure it has not been switched off or buried in the configuration.
• The European client imposes its own chatbot: the centre operates on the client's behalf and must flag any configuration that would remove the information.

The exact classification depends on the facts; our guide to the provider or deployer role sets out the method. In every case, the contract must state who drafts, approves and keeps the disclosure text.

When the "it is obvious" exception does not apply#

The obligation falls away if a reasonably well-informed and observant person understands unaided that they are interacting with AI. The Commission asks for this exception to be read narrowly. In a contact centre, it almost never applies:

• today's synthetic voices sound close to a human voice;
• giving the bot a first name ("Sarah") suggests, on the contrary, a human;
• a customer stressed by an outage or a dispute pays no attention to cues;
• the conversation can switch between AI and an agent without the customer noticing.

The prudent rule is simple: inform explicitly, in every new conversation.

Example wording for chat and voice#

The Commission's final guidelines, published on 20 July 2026, specify that customer-service or sales AI agents must state that they are AI and on whose behalf they act, from the moment of contact. The texts below are examples to adapt, not official templates. They are given in French, as used with French customers, with an English rendering in parentheses.

Chat, first message: « Bonjour, je suis l'assistant virtuel de [Marque], un programme d'intelligence artificielle. Je peux répondre à vos questions sur votre contrat. Tapez « conseiller » à tout moment pour parler à une personne. » ("Hello, I am [Brand]'s virtual assistant, an artificial intelligence programme. I can answer your questions about your contract. Type 'adviser' at any time to speak to a person.")

Interactive voice response, welcome message: « Bienvenue chez [Marque]. Vous échangez avec un assistant vocal automatisé, fonctionnant par intelligence artificielle. Pour être mis en relation avec un conseiller, dites « conseiller » ou tapez zéro. » ("Welcome to [Brand]. You are speaking with an automated voice assistant powered by artificial intelligence. To be put through to an adviser, say 'adviser' or press zero.")

Outbound voicebot, first sentence: « Bonjour, ici l'assistant vocal automatisé de [Marque]. Je suis une intelligence artificielle et je vous appelle au sujet de votre rendez-vous du [date]. » ("Hello, this is [Brand]'s automated voice assistant. I am an artificial intelligence and I am calling you about your appointment on [date].")

Handover to a human: « Je vous transfère à un conseiller de [Marque]. » ("I am transferring you to a [Brand] adviser.") Then, on the agent's side: « Bonjour, [prénom], conseiller chez [Marque], je reprends votre demande. » ("Hello, this is [first name], adviser at [Brand]; I am taking over your request.")

A few drafting rules:

1. Place the information before any question to the customer, not after.
2. Use the words "artificial intelligence" or "automated", not just "assistant".
3. Name the brand on whose behalf the agent acts, not the Moroccan service provider if the customer does not know it.
4. Offer a way out to a human, which also serves the customer experience.
5. Write the message in the language of the conversation and make it accessible (text readable by screen readers, slow speech rate, no music drowning out the voice).

An outbound voicebot used to prospect consumers in France must also comply with the French rules on telemarketing (Law No. 2025-594), which have required prior consent since 11 August 2026.

Synthetic voices and generated content#

If the voicebot produces a synthetic voice or the centre generates audio messages, Article 50(2) requires the provider of the generative model to apply machine-readable marking. A code of practice on the transparency of AI-generated content was finalised on 10 June 2026 and recognised by the Commission and the AI Board as an adequate voluntary tool in July 2026. The centre should ask its software vendor how this marking is ensured.

A voice that imitates a real person (an executive, a spokesperson) also falls under the deepfake regime of Article 50(4): it must be disclosed as artificial.

Analysing a customer's emotions from their voice, for its part, triggers the information duty in Article 50(3); see our guide to emotion analysis in call centres.

What you must be able to prove#

In an inspection or a client audit, the statement "we inform customers" is not enough. A centre's evidence file includes:

• the dated version of each welcome script (chat, IVR, outbound) and its change history;
• screenshots and test recordings showing the information at the first exchange;
• conversation logs proving that the message is triggered in every session;
• the contractual allocation of responsibilities with the software vendor and the client;
• a periodic review, at least whenever the bot is updated.

These items also feed into answers to European clients' compliance questionnaires, which increasingly ask about AI transparency.

What a breach costs#

A breach of Article 50 falls within the fine bracket of Article 99(4): up to €15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs and start-ups, it is the lower of the two. For a Moroccan centre, the immediate risk is above all contractual: a client can suspend a non-compliant set-up or use it as grounds for not renewing.

Disclosing AI is also part of service quality: a customer who knows who they are talking to and can reach a human accepts automation more readily, which ties in with measuring quality in a contact centre.

Article 50 checklist for a contact centre#

1. List every automated touchpoint: web chat, messaging, IVR, inbound and outbound voicebots, generated emails.
2. Identify those that serve people in the European Union.
3. Check that each touchpoint informs from the first exchange, using the words "artificial intelligence" or "automated".
4. Add the brand on whose behalf the agent acts.
5. Provide a tested way out to a human agent.
6. Check the accessibility of the message (written and spoken).
7. Obtain from the software vendor an explanation of how synthetic content is marked.
8. Archive scripts, tests and logs, and schedule a review at every update.
9. Write these points into contracts with clients and software vendors.