Money · Operator memo

Does Article 50 apply to my AI voice agent or phone bot?

Yes if the voice agent itself talks to a person. The Commission FAQ lists chatbots, AI agents, and avatars as systems that interact directly with people. A realistic human-sounding phone voice is not “obvious.” The obviousness exception is to be read restrictively.

Article 50(1) notice has been enforceable since 2 August 2026. Synthetic audio also has an Article 50(2) machine-readable mark. If you clone a real person’s voice, that is deepfake audio: deployers need an audible label, not just the hidden mark.

Get the $29 Article 50 Ship Checklist Buy a Money Brief $149

Not legal advice. Related: Email support agents · Deepfake labels · Free notice tool

The four FAQ criteria, applied to voice

The Article 50 FAQ (last update 24 July 2026) lists four cumulative criteria for the interaction-notice duty. All four have to hold:

  1. The system is an AI system.
  2. It is designed for a genuine two-way exchange with people, not merely collecting data or providing automated responses.
  3. The interaction is direct: the AI itself communicates with the person, not through a human intermediary.
  4. The person is a natural person (consumer, professional, or other user).

A conversational phone or in-app voice agent (speech-to-text, a model, text-to-speech back to the caller) usually clears all four. A DTMF menu that only says “press 1 for sales” is closer to the FAQ’s automated-responses carve-out. Criterion (3) is still the fork: who is speaking.

Who is speaking

What you shipped Art. 50(1) notice? Why
Conversational AI answers the call and talks to the person Yes Direct two-way AI-to-person interaction. FAQ examples include AI agents.
Human-sounding cloned voice, no spoken “this is AI” Yes, and treat “obvious” as out FAQ: identify a reasonably well-informed, observant person. A voice that sounds human is the opposite of obvious.
AI drafts; a human takes the call or repeats the answer Usually no under Art. 50(1) FAQ: not direct when it goes through a human intermediary.
Old-style IVR: “press 1”, no conversation Usually no FAQ carves out systems that only provide automated responses / collect data.
US company, EU callers or EU-used output Still in FAQ: third-country providers are in if the output of the system is used in the EU.

People must be informed from the start of the first interaction, in a clear and distinguishable way, including accessibility. On a phone line that is an audible line before the conversation continues, not a later SMS or a privacy policy.

Two more duties for voice, not just the notice

Article 50(2): providers of systems that generate synthetic audio must mark that audio in a machine-readable format so it can be detected as AI-generated. A limited grace period exists only for systems already on the market before 2 August 2026: those marks are due 2 December 2026. The spoken notice is a different duty and had no extra delay.

Article 50(4) deepfakes: AI-generated or manipulated audio that resembles an existing person and would falsely appear authentic is a deepfake. Deployers must disclose it to the person upon first exposure, in a clear and distinguishable manner, perceivable without technical tools. The FAQ’s example is visible or audible labels. The hidden Article 50(2) mark does not satisfy this.

A cloned founder or agent voice on outbound sales or support calls is the case that trips both the interaction notice and the deepfake audio label.

What is live this week

Article 50 applies from 2 August 2026. The 31 July 2026 Commission press says chatbots and other interactive AI systems have to tell users they are dealing with AI, not a human, and that deepfake audio has to be labelled. The Digital Omnibus moved high-risk dates, not Article 50.

National market surveillance authorities enforce most Article 50 cases. Fines can reach €15 million or 3% of worldwide turnover, with proportionality language for SMEs.

Do this in seven days

  1. Inventory every surface where a model speaks: inbound phone, outbound dialler, in-app voice, voicemail drops, WhatsApp voice notes that auto-play.
  2. If the AI talks to the person, ship an audible first-interaction line. Commission-style example used on this desk: “You are interacting with an AI system.”
  3. Do not treat a human-sounding TTS voice as “obvious.” The FAQ says read that exception restrictively.
  4. If you clone a real person’s voice, add an audible deepfake label as well as the spoken notice. The machine-readable mark is not enough.
  5. Open an Article 50(2) marking ticket for the audio pipeline. Pre-2 August products get until 2 December for that mark only.

Buy the checklist

The $29 Article 50 Ship Checklist is the print-and-tick page: inventory, notice line, deepfake labels, 2 December mark. The $149 Money Brief is the sourced memo if your stack mixes voice agents, cloned voices, and EU callers and legal wants the FAQ criteria applied line by line.

Get the $29 Article 50 Ship Checklist Buy a Money Brief $149

Free notice-copy tool · Omnibus did not delay Article 50

Sources

Fetched 28 August 2026 (PT). No invented quotes.

  1. European Commission, “Transparency obligations under Article 50 of the AI Act,” FAQ, last update 24 July 2026 (chatbots, AI agents, and avatars as direct-interaction examples; four Art. 50(1) criteria including human intermediary and automated-responses carve-out; obviousness exception read restrictively; third-country providers in if output is used in the EU; Art. 50(2) machine-readable marks for synthetic audio; Art. 50(4) deepfake audio with audible labels; applies from 2 August 2026; 2 December grace only for Art. 50(2) marks; fines up to €15m / 3%). digital-strategy.ec.europa.eu
  2. European Commission, “Commission starts enforcing AI Act rules and new transparency requirements on 2 August,” press release, 31 July 2026 (chatbots and other interactive AI; deepfake audio labelled; machine-readable marks). digital-strategy.ec.europa.eu
  3. European Commission, AI Act policy page, last update 3 August 2026. digital-strategy.ec.europa.eu