EU AI Act Article 50: what AI voice agents must tell callers

Why the AI Act makes voice agents say they are AI

The EU AI Act (Regulation (EU) 2024/1689) sorts AI systems by risk. A few practices are banned, high-risk uses such as credit scoring carry heavy duties, and most systems carry few. Article 50, the transparency article, cuts across that scale. It applies to systems that interact with people or generate content, whether or not they are high-risk.

The reason is set out in the Act's recitals. Some AI systems can create risks of impersonation or deception even when they are not high-risk, so people "should be notified that they are interacting with an AI system, unless this is obvious" (recital 132). The same recital asks providers to take account of people who are vulnerable because of their age or disability, and to give notices in accessible formats. A voice agent that sounds human is close to the centre of that concern.

The core duty for a voice agent is Article 50(1): people must be told they are talking to an AI system, unless that is obvious. Three more duties can apply to a calling operation. Synthetic speech must carry a machine-readable mark (Art. 50(2)). People must be told when a system detects their emotions (Art. 50(3)). And a voice that imitates a real person must be disclosed as AI-generated (Art. 50(4)).

Article 50(5) then sets the manner for all of them: clear, distinguishable and accessible, at the latest at the first interaction. The duty is the same for a text chatbot and a voice agent; what differs is how the notice reaches the person. The timing has not moved either. The Digital Omnibus (Regulation (EU) 2026/1744) delayed the high-risk rules but not Article 50, and only machine-readable marking for systems already on the market got extra time.

Providers, deployers and companies outside the EU

Because the duties fall on different parties, the first practical step is to work out your role. The provider builds the AI system, or has it built, and offers it in the EU under its own name (Art. 3(3)). The deployer is the business that uses the system under its authority (Art. 3(4)). If a bank buys a ready-made calling agent and runs campaigns with it, the bank is the deployer and the company selling the agent is the provider.

The roles can merge. According to the Commission's guidelines on Article 50, a company that builds its own agent in-house and uses it under its own name is a provider. So is a company that modifies someone else's system and offers it under its own name.

The split matters because Article 50(1) puts the duty on the provider: the agent must be built so that callers are informed. In practice, though, the deployer writes the script and picks the voice, and a deployer who edits the greeting can remove the disclosure without the provider knowing. Both sides should agree in the contract who makes sure the disclosure is spoken on every call, and the deployer should check its live prompts and opening line itself, in every language it calls in.

What to say, and when: the first sentence of the call

The timing is strict. An AI phone agent has to say it is an AI at the latest at the first interaction, which on a call means the opening line (Art. 50(5)). The guidelines give a telephony example: an explicit spoken statement at the start, such as "This is an AI-powered assistant". Tones or other audio cues may help, but they are not enough on their own.

The guidelines add three points for calls. A single clear notice at the start usually suffices. In riskier contexts, such as financial advice, insurance, health advice, complaints, or calls with older or vulnerable people, periodic reminders are likely to be needed. And the agent must disclose that it is an AI whenever someone asks about its nature. AI agents must also state on whose behalf they are acting, so the opening line should name the company, not just the fact that the caller is an AI.

Our advice on how to deliver it goes beyond the text. Play the disclosure as fixed, pre-written words rather than a line the language model generates, so it cannot be skipped or paraphrased away, and stop callers from talking over it. Developers building on open-source voice frameworks describe the same pattern. Write the line in each language you call in, in short, plain words, and test it with real listeners, including older callers, because Article 50(5) requires the notice to meet the applicable accessibility rules.

If you record the call, the same opening can carry the recording notice the GDPR requires, covered below. A line such as "Hello, this is an AI-powered assistant calling on behalf of [Company]. This call is recorded; you can find our privacy notice at [link]" does both jobs. The AI disclosure and recording-notice scripts have versions in several languages.

Finally, keep evidence that the disclosure played. On Sonex, for example, every call on an outbound calling campaign is kept as a recording and a transcript, which shows what was said in the first seconds if a regulator or customer asks.

Why the obvious-use exception rarely helps on calls

It is rarely obvious enough on a phone call that the caller is talking to an AI. Article 50(1) drops the duty only when AI use is obvious to a person who is reasonably well-informed, observant and circumspect, given the context. The guidelines say this exception must be read narrowly and kept to cases where almost no doubt is left.

Several factors work against relying on it. The guidelines list a human-sounding voice as something that makes AI use less obvious. If the general public can reach the system, including children, older people or people with disabilities, the exception cannot be relied on for them. Their examples of obvious cases are code assistants for developers or internal tools for trained staff, not customer calls.

Accessibility is part of the duty too. Article 50(5) requires the information to meet the applicable accessibility rules, and the guidelines point to the European Accessibility Act (Directive 2019/882). They do not add accessibility requirements of their own beyond those rules. Short, plain wording in the caller's own language is the safest way to meet them.

Machine-readable marking and cloned voices

The spoken disclosure is for the listener; Article 50(2) adds a technical duty for the audio itself, and it is the provider's job. Providers of AI systems that generate synthetic audio must make sure the output is marked in a machine-readable format and detectable as AI-generated (Art. 50(2)). The guidelines confirm that audio includes speech.

The Commission's Code of Practice on marking and labelling AI-generated content, published on 10 June 2026, describes how. Signatories commit to at least two layers of machine-readable marking, such as metadata and an imperceptible watermark, and to provide a way to detect the marks. On 8 July 2026 the Commission found the Code adequate for Articles 50(2), (4) and (5). Signing is voluntary, but it is one way to show compliance, so ask your voice provider how its output is marked and whether it has signed.

The deadline depends on when the system reached the market. Systems already on the market before 2 August 2026 must take the steps needed to comply with Article 50(2) by 2 December 2026 (Art. 111(4), added by the Digital Omnibus). Systems placed on the market from 2 August 2026 onwards have had to comply from the start.

A cloned or human-like voice brings in the deepfake rule as well. A voice counts as a deepfake when it resembles an existing person and would falsely appear authentic (Art. 3(60)). Cloning the voice of a real employee, founder or celebrity for a calling agent is the clear case. The deployer must then disclose that the audio is AI-generated (Art. 50(4)), and should hold the person's consent to use their voice.

The guidelines also treat realistic AI-generated personas and avatars as "persons" for this test (para 113). So, on our reading, it could be argued that a named, human-sounding agent that never says it is an AI falls in as well. A clear AI disclosure at the start of the call addresses both Article 50(1) and this risk.

Emotion recognition in call centres

Article 50(3) matters if your system reads emotion from people's voices, and here the AI Act draws a sharp line between customers and staff. AI systems that infer the emotions of people at work or in education have been banned since 2 February 2025, except for medical or safety reasons (Art. 5(1)(f)). The Commission's guidelines on prohibited practices give the exact example: a call centre using webcams and voice recognition to track its employees' emotions, such as anger, is prohibited.

Tracking customers' emotions, such as anger or impatience, is not banned by Article 5(1)(f), the same guidelines say. Their example is a system that helps employees cope with angry customers. But the deployer must tell the people exposed to it (Art. 50(3)), and emotion recognition is a high-risk use whose extra duties apply from 2 December 2027. Under the AI Act, emotion recognition by definition works on biometric data such as the voice (Art. 3(39)). Whether the same processing is special-category biometric data under the GDPR depends on whether it identifies the person, as explained in when a voice becomes biometric data.

• Your staff, from voice or other biometric data: banned since 2 February 2025, unless for medical or safety reasons.
• Customers, from voice or other biometric data: allowed, but tell them (Art. 50(3)); high-risk duties apply from 2 December 2027.
• Anyone, from the text transcript only: not emotion recognition under the AI Act, but the GDPR still applies, and labour law too for staff.

Fines, and how Article 50 fits with the GDPR

Breaching Article 50 can bring a fine of up to €15 million or 3% of worldwide annual turnover, whichever is higher (Art. 99(4)(g)). For small and medium-sized enterprises, including start-ups, the lower of the two applies (Art. 99(6)). The Digital Omnibus extended that rule to small mid-cap companies (Art. 99(6a)).

Breaching the Article 5 bans costs more. Using emotion recognition on your own staff, for example, can bring a fine of up to €35 million or 7% of worldwide annual turnover (Art. 99(3)).

Article 50 also sits alongside the GDPR rather than replacing it: they are separate duties, and you need both. Article 50 is about telling people they are dealing with an AI. The General Data Protection Regulation (GDPR) is about personal data, and a call recording is personal data. The AI Act says it does not affect the GDPR (Art. 2(7)).

Under Article 13 GDPR, you must tell people, when you collect their data, who you are, why you process it and on what legal basis. For a recorded AI call, the opening line usually covers both the AI and the recording, with a route to the full privacy notice. Some EU countries also have national rules on recording calls. We have not checked them country by country, so check each country you call; the calling rules by country tool is a starting point.

Key dates