EU AI Act Article 50: what your voice agent must actually say

Most AI phone agents open with some version of "Hi, I'm Ava, your virtual assistant." Since 2 August 2026 that sentence has been a compliance artefact in the European Union, and the Commission's own guidance says the word "assistant" on its own does not do the job.

The obligation is short. The guidance explaining how to meet it runs to 51 pages, and the operational detail — what a voice agent must say, when it must say it again, and what it must do when a caller asks whether it is human — sits in three paragraphs almost nobody has read.

What changed on 2 August 2026, and what 2 December is not

Article 50 of the AI Act has applied since 2 August 2026, per the Commission's transparency FAQ. The obligation itself reads: "Providers shall ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system" (Article 50(1), as reproduced by the AI Act Explorer).

A second date is circulating in vendor newsletters and being read as a grace period. It is not. Per the same FAQ, the transition to 2 December 2026 applies only to the Article 50(2) marking duty for synthetic content, and only for systems already placed on the market before 2 August 2026. If you switched on a voice agent this year, your disclosure duty started in August.

The penalty band is set by Article 99(4): up to EUR 15,000,000 or 3% of worldwide annual turnover for breaches of Article 50.

The telephony paragraph everyone skipped

On 20 July 2026 the Commission adopted guidelines on these transparency obligations — document C(2026) 5054 final, whose annex runs to 51 pages (Commission library entry, full PDF). Paragraph 37 addresses voice directly:

"Auditory disclosure: In voice-based or telephony contexts, explicit spoken statements at the beginning of the interaction (e.g. 'This is an AI-powered assistant') combined, as appropriate, with periodic reminders in longer interactions, in particular in case of interruptions or a change of the role of the AI system during a user journey."

Three design requirements are hiding in one sentence: spoken, at the beginning, and repeated in long or interrupted calls. The same paragraph closes a popular shortcut — distinctive audio cues or earcons "are not considered sufficient by themselves".

Four things the Commission says do not work on their own

Paragraph 38 lists disclosure methods that fail when used alone. Burying it in terms and conditions, a URL or product documentation. Machine-readable markings that are not perceivable at the point of interaction. And the one that catches most voice deployments: "Unclear or ambiguous signals (e.g. generic references to 'assistant')".

Paragraph 142 puts the test plainly — a disclosure is not clear and distinguishable if it can be "easily overlooked" or is "part of terms of use that are often not read by users".

There is an exception for cases where the AI nature is obvious to a reasonably well-informed person, but the guidelines' examples draw the line where it hurts: it does not cover chatbots embedded in online platforms or helpdesk support tools, and it does cover internal employee-facing assistants. A customer phone line is on the wrong side of that line.

When once is not enough

Paragraph 40 says a single prominent notification is usually sufficient, then names the contexts where periodic reminders are "likely to be necessary" because users may be misled: financial advice, insurance, legal assistance, health advice, and complaints handling.

Read that list against a real contact centre. A complaint escalation that starts as an order query is exactly the "change of the role of the AI system during a user journey" that paragraph 37 flags. The design implication is that disclosure is a per-turn behaviour, not a greeting.

"Am I talking to a human?" — the answer you cannot design around

The same paragraph turns the most common caller question into a build requirement. Providers "must also design and develop the AI system to ensure disclosure in all situations where the AI system is being asked questions relating to its nature or to the origin of the interaction."

That is not a policy for agents to follow. It is an obligation on the system's design: the agent must answer honestly when asked, every time, including when a caller phrases it as "you're a real person, right?" Any prompt-level instruction that can be talked around is a system that fails this test on some calls — and a sampled QA programme will not tell you which ones.

Every new call is a new first interaction

Paragraph 143 settles the question of how often you say it. The information "should, at least, be provided once at the start of an interactive session", and the first-interaction concept covers "any subsequent, first interaction… by any other natural person".

For a phone line, that means every call. There is no concept of a returning caller who has already been told. If a customer rings three times in a week, the agent discloses three times.

The US bar exists, and it is lower

Teams operating on both sides of the Atlantic should not assume US compliance covers them. In FCC 24-17, adopted 2 February 2024, the Commission confirmed that the TCPA's restrictions on an "artificial or prerecorded voice" encompass current AI technologies that generate human voices. But the identification duty at 47 CFR 64.1200(b)(1) is about identifying the caller at the start of the call, not about disclosing that the voice is synthetic.

State law goes further in places. California's Business and Professions Code §17941 requires bot disclosure that is "clear, conspicuous, and reasonably designed to inform", and Public Utilities Code §2874(a)(3) requires informing the person called if a prerecorded message uses an artificial voice. Maine added a disclosure duty at 10 MRSA §1500-Y through PL 2025 c.294, approved on 12 June 2025, enforced through its Unfair Trade Practices Act.

None of these carries the EU's repeat-disclosure or answer-when-asked duties. Build to the higher bar and the lower one is covered.

Latency turns into a compliance problem

Paragraph 37's "in case of interruptions" clause matters because interruptions are common. Full-Duplex-Bench-v3, posted on 6 April 2026, measured live voice systems and reported a 13.5% interruption rate for GPT-Realtime at Pass@1 of 0.600, response latency as low as 4.25 seconds for the fastest system tested, and 10.12 seconds for a cascaded speech-to-text, LLM and text-to-speech pipeline. A slow agent gets talked over, and an agent that gets talked over is one whose opening disclosure may never have landed.

Eight questions for a voice vendor, before you sign

Every one of these maps to a sentence quoted above, so the answers are checkable rather than a matter of opinion.

• Is the opening disclosure spoken, and is it a pre-rendered asset verified before the call connects, or synthesised at runtime?
• What exactly does it say? If the only signal is the word "assistant", paragraph 38 calls that ambiguous.
• Does it fire on every call, including repeat callers, as paragraph 143 requires?
• What triggers a repeat disclosure — call duration, an interruption, or a change in the call's purpose?
• Is complaints handling treated as a context needing reminders, as paragraph 40 names it?
• When a caller asks whether they are talking to a person, is the honest answer enforced by the system, or by a line in a prompt that a determined caller can talk around?
• If the caller interrupts the opening, does the agent complete the disclosure on its next turn, or drop it?
• What is logged, per call, to evidence all of the above six months later?

A vendor who answers "it's in our terms of service" has answered paragraph 38 for you, in the negative.

What evidence survives a market-surveillance request

These obligations are per call and per turn. Sampling two percent of calls cannot evidence that every caller heard the disclosure, that reminders fired in every escalated complaint, or that the agent answered honestly every time it was asked. Four artefacts are worth having before anyone asks for them: a pre-rendered opening disclosure asset verified at the start of each call, a log of every reminder with its trigger, a transcript search proving the answer given each time nature was questioned, and a dated record of the script version in force.

VoiceSieve is being built to that shape: the agent states it is an AI in its first sentence, repeats it when the call changes character, answers plainly when asked, and hands to a person the moment the answer is not in an authoritative source. The design predates the guidelines. The guidelines simply made it the minimum.