AI voice agent disclosure: telling callers they are talking to AI

What changed for AI voice agents on 2 August 2026?

Article 50 of the EU AI Act, Regulation (EU) 2024/1689, began to apply on 2 August 2026. Its first paragraph turns AI voice agent disclosure into a design requirement: a system intended to talk directly with people must let them know they are dealing with AI, unless a reasonable person would find that obvious. The same article covers marking synthetic audio, emotion recognition notices and deep fake labels.

On 20 July 2026 the European Commission published guidelines on Article 50, adopted as C(2026) 5054. They are not binding, but they set out the Commission's reading and list voice assistants, customer support chatbots and AI hotlines among the examples in scope. A consolidated text dated 27 July 2026 includes the amendments made by Regulation (EU) 2026/1744, among them the transitional marking rule the guidelines attribute to the AI Omnibus.

This article is general information for operations and customer experience teams, not legal advice. Confirm how the rules apply to your systems with your legal owner.

Why does disclosure matter for contact centres?

Phone channels are where the obviousness exception is weakest. Callers cannot see an interface, synthetic voices can sound human, and a public line reaches everyone, including older callers and people with disabilities. The Commission links the duty to reducing impersonation, deception and fraud, and to helping people judge how far to trust what they hear.

The penalties are material. Under Article 99(4), breaching Article 50 can lead to fines of up to EUR 15 million or, for an undertaking, up to 3% of worldwide annual turnover in the previous financial year, if that is higher; for SMEs the lower figure applies. National market surveillance authorities do most of the enforcement, and anyone with grounds to suspect a breach can complain to them.

AI voice agent disclosure: provider or deployer?

Article 50(1) puts the disclosure duty on the provider: whoever develops the voice agent, or has it developed, and places it on the market or puts it into service under its own name. A contact centre using a vendor's agent under its authority is a deployer, and the deployer still has work to do.

Many enterprises sit on both sides. The definition in Article 3(3) covers having a system developed by someone else, and the guidelines treat an organisation that builds a chatbot in house and runs it under its own name as its provider. The same applies to a company that modifies an existing generative system and puts it into service under its own name.

Deployers have duties of their own. If a feature is an emotion recognition system, meaning it infers emotions or intentions from biometric data, Article 50(3) requires the deployer to tell callers it is operating, and the guidelines note that such systems are also classified as high-risk. A voice that resembles a real person and could pass as authentic may be a deep fake the deployer must disclose under Article 50(4). Separately, Article 5(1)(f) prohibits using AI to infer the emotions of staff at work, except for medical or safety reasons.

When is it obvious that a caller is talking to AI?

Rarely, on a customer line. Article 50(1) lifts the duty only when the AI nature is obvious to a person who is “reasonably well-informed, observant and circumspect”, given the circumstances and the context of use. The Commission says the exception must be read restrictively and should cover only cases where almost no doubt remains for an average member of the audience.

The test starts with who is likely to call. A public line reaches children, older people, people with disabilities and people with little AI literacy, and the guidelines say the exception cannot be relied on to protect them. A human-sounding voice makes the AI nature less obvious, and general awareness that AI agents exist does not count.

The exception can work in narrow internal settings, such as the guidelines' example of an assistant used by trained, AI-literate staff for HR or IT support. Record the assessment either way; the provider has to be able to demonstrate it.

How do you disclose on a call without slowing it down?

Put one clear sentence in the greeting, before the caller shares anything, and make it name AI. Article 50(5) requires the information to reach people “in a clear and distinguishable manner” no later than the first interaction, and to meet the accessibility requirements that apply. For telephony, the guidelines suggest an explicit spoken statement at the start, with reminders in longer calls.

Some shortcuts do not count on their own: a notice only in terms and conditions or on a web page, marks the caller cannot perceive, a generic reference to an assistant, or a statement about the technology alone. A tone or earcon can help blind and partially sighted callers but is not sufficient by itself.

Speed is a design problem, not a legal one. Combine the notice with the purpose of the call in one sentence, and repeat it in the next turn if the caller talks over the greeting. The guidelines accept that one prominent notice will usually be enough.

Repeat it when it matters. The guidelines say the system must disclose its nature whenever a caller asks, and expect reminders in riskier contexts such as complaints, financial advice, insurance and health. An agent acting on someone's behalf, for example on an outbound call, should also say whose.

What happens when the call moves to a human agent?

Once a person takes over and speaks for themselves, the caller is no longer interacting directly with the AI system, so Article 50(1) does not apply to that part of the call, nor to staff who use AI tools to help them talk to customers. The duty returns whenever AI-generated speech reaches the caller again.

Blended journeys need care. AI-generated replies mixed into a human conversation still need disclosure unless a person has properly reviewed them and is the main interlocutor, and the guidelines call for reminders when the role of the AI system changes during a journey. A handback to the agent for payment capture or a survey is exactly that moment.

Make the handover audible both ways and pass the context so the caller does not repeat themselves. Article 50 does not require you to offer a person; that is a service decision, and often the simplest answer for callers who would rather not talk to a machine.

Recordings, transcripts and synthetic voices

Article 50 does not decide whether you may record a call; Article 50(6) and the guidelines leave other transparency and data protection duties in place. What Article 50(2) adds is a provider duty to mark synthetic audio, image, video and text in a machine-readable form and make it detectable as AI-generated.

The guidelines suggest that live content consumed immediately, and never recorded, stored or passed on, may be exempt where marking is not technically feasible and people are told it is AI-generated. A recorded call does not fit that description, which leaves three questions for your provider.

The transitional rule is narrow. Article 111(4), added by Regulation (EU) 2026/1744, gives providers of generative systems placed on the market before 2 August 2026 until 2 December 2026 to meet Article 50(2). For a system that is both interactive and generative, the Commission says this covers marking only. Content generated before 2 August 2026 need not be marked retroactively.

• Is the agent's speech marked in stored recordings?. Ask how AI-generated audio is marked and how the marks are detected. The guidelines treat marking without detection as incomplete.
• Are post-call summaries covered?. The guidelines list AI-generated summaries of text among outputs that need marking, while plain transcripts count as standard editing.
• Does the voice resemble a real person?. Realistic speech in a specific person's voice is listed as needing marking, and a convincing cloned voice may also be a deep fake the deployer must disclose.

Worked example: an energy retailer's billing line

Picture an energy retailer that licenses a voice agent from a vendor for inbound billing calls and outbound payment reminders. The agent uses a stock synthetic voice, calls are recorded for quality review, and the platform offers an optional sentiment score.

• Map the roles. The contract confirms the spoken AI notice is built in, cannot be switched off and is retested after each release. Because the retailer runs the agent under its own brand, it asks its legal owner whether that also makes it a provider.
• Write the openings. Inbound calls open with the AI notice and the agent's scope. Outbound reminders say an AI agent is calling on behalf of the retailer before any security questions.
• Plan the handovers. Disputes and signs of distress go to a person with the call context. When the call returns to the agent for card payment, the agent says so first.
• Check the extras. The team leaves the sentiment score off until it has checked whether it is an emotion recognition system under Article 3(39), and asks the vendor how recorded synthetic audio is marked.

Common misconceptions about AI disclosure on calls

Most mistakes come from treating disclosure as a legal notice rather than part of the conversation.

• A recording notice is enough. Telling callers that calls may be recorded says nothing about AI. The caller needs to hear that the voice itself is an AI agent.
• Calling it a virtual assistant is enough. The guidelines name generic references to an assistant as unclear. Say AI, or AI agent, in plain words.
• The extension buys time for the greeting. The 2 December 2026 deadline covers only machine-readable marking under Article 50(2) for older systems. Spoken disclosure has applied since 2 August 2026.
• Article 50 requires a human option. It does not. A route to a person is a service choice, although the guidelines expect extra care in riskier contexts such as complaints.

A readiness checklist for CX leaders

Run through this list for each AI voice journey before and after every release, and keep the results as evidence.

• Inventory. List every line, campaign and menu path where AI speech reaches a caller, including overflow routes.
• Roles. Record who is provider and who is deployer for each agent, why the obviousness exception does or does not apply, and who owns disclosure under the contract.
• Opening and reminders. Confirm the AI notice plays before the caller gives any information, survives barge-in and exists in every supported language. Script answers for handbacks, complaints and questions about whether the agent is human.
• Accessibility. Check pace, clarity and alternative channels against the accessibility rules that apply to your service.
• Synthetic content and deployer duties. Ask how audio, summaries and any cloned voices are marked, and check emotion and biometric features against Article 50(3).
• Evidence and training. Retest after every model, voice or prompt change and brief the staff who take transfers. Article 4 asks providers and deployers to support AI literacy among the people who operate their systems.

How Bibha supports the notice process

Bibha's Realtime voice capabilities include disclosure and consent controls for applying the notice and recording-consent process your team approves, plus human transfer with relevant context and permitted transcripts and recordings. PeachDesk, a voice-agent solution built on Bibha, covers inbound and outbound call agents, voice configuration, telephony, outbound campaigns and call analytics.

These capabilities help you implement agreed controls; they do not by themselves establish compliance. Whether a configuration meets Article 50 depends on the actual system, scripts and operating practice.

Questions and answers

References

1. European Commission. (2026, 20 July). Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689 (the 'AI Act') (C(2026) 5054 final). https://ai-act-service-desk.ec.europa.eu/sites/default/files/2026-07/guidelines_on_the_implementation_of_the_transparency_obligations_for_certain_ai_systems_under_article_50_of_the_ai_act_bzptwqhk0ikg1dtlddap41psfy_131215.pdf (external site)
2. European Commission. (2026, 20 July). Commission publishes guidelines on transparency obligations for providers and deployers of certain AI systems. Shaping Europe's digital future. https://digital-strategy.ec.europa.eu/en/news/commission-publishes-guidelines-transparency-obligations-providers-and-deployers-certain-ai-systems (external site)
3. European Commission. (2026, 24 July). Transparency obligations under Article 50 of the AI Act [Questions and answers]. Shaping Europe's digital future. https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act (external site)
4. European Commission. (2026). Code of Practice on Transparency of AI-Generated Content. Shaping Europe's digital future. https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content (external site)
5. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). (2024). Official Journal of the European Union, L, 2024/1689. Consolidated text of 27 July 2026, including Regulation (EU) 2026/1744. EUR-Lex. https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng (external site)