AI agent experience

Your AI agent has to say what it is.

EU AI Act, Article 50 · applies 2 August 2026

From the 2nd of August 2026, EU law requires AI systems that deal directly with people to make that clear to them. The duty is Article 50 of the AI Act, it has not moved, and the easiest place to fail it is a phone line. Here are twelve checks you can run against your own AI agent today, with your own phone.

What the law says

"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, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use." Article 50(1), Regulation (EU) 2024/1689 · EUR-Lex

Four things worth knowing before you run the check:

  1. The duty sits with the provider of the system. If the AI agent runs under your name and your brand, assume that is you: the Act treats putting your own name on a vendor's system as taking on the provider's obligations.
  2. On the Commission's guidelines, "obvious from context" will not save a customer-service AI agent: the guidelines name helpdesk chatbots as exactly a case where AI is not obvious, and say the exception is to be read restrictively. A human-sounding voice only makes it less obvious.
  3. The disclosure has to land clearly and distinguishably, at the latest at first contact. The guidelines are explicit that it has to be given at least once at the start of every session and to every new caller, that riskier contexts need reminding again, and that a line in your terms and conditions does not comply.
  4. Non-compliance carries fines up to 15 million euros or 3 percent of total worldwide annual turnover, whichever is higher; for SMEs, the lower. Enforcement is national and still uneven; the obligation applies regardless.

The twelve-point check

Is the disclosure there?

  • A1 The voice greeting says it is an AI before anything else happens. Before the menu, before the small talk, before the identity questions.
  • A2 The chat says it is an AI at the first message. In the conversation itself, not behind an info icon or a help page.
  • A3 Outbound calls disclose at pickup. Not partway through the script, and not only if the customer asks.
  • A4 Every entry path carries it. IVR deflection, callback, after-hours cover, the transfer out of a human queue.

Would a caller actually catch it?

  • B5 It is spoken plainly, at talking speed. Not rushed, and not folded into the recording-and-privacy preamble.
  • B6 It survives interruption. Callers barge in. If the disclosure dies with the greeting, it was never heard.
  • B7 It is in the language of the conversation. Not only in the language of head office.
  • B8 The persona does not fight the disclosure. A human name, breath sounds and "let me just check that for you" can un-say what the greeting said.

Does it hold at the edges?

  • C9 Asked "am I talking to a human?", the AI agent answers truthfully. Every phrasing, every time, however friendly the persona.
  • C10 Handover is marked. When a person takes over, the caller can tell the human has started.
  • C11 Your vendor confirms the call audio carries machine-readable marking. Article 50(2) expects synthetic audio to be detectable as such. How that applies to live calls is not yet settled, and systems already on the market before the 2nd of August 2026 have until the 2nd of December 2026 for this marking; ask the question and keep the answer in writing.
  • C12 If the voice is cloned from a real person, the deepfake rules apply on top. Article 50(4). A generic branded voice is not a deepfake; a cloned one is.

Rows A1 to A4, C9, C11 and C12 follow from the Article; where its guidance is silent, they are its plain reading. The rest is what it takes to pass them with a real caller on the line.

The three-call self-test

Ten minutes with your own phone.

  1. Call 01 Say nothing. Listen to the whole greeting. Is the disclosure there, and would a first-time caller catch it?
  2. Call 02 Interrupt the greeting straight away, the way real callers do, and ask your question. Does the disclosure still happen?
  3. Call 03 Mid-conversation, ask "am I speaking with a human?". Score the answer word for word.

Score honestly. A miss on A1 or C9 is a miss on the Article. A miss on B6 means a real caller can hang up never having been told, which is the outcome the Article exists to prevent, whatever the vendor dashboard says.

The free probe

I will run the three calls for you.

For any company whose AI agent answers calls from customers in the EU: I am Dylan Evans. I have run EU-regulation-driven programmes at Philips, and I have directed system voices since 2001: both halves of what this probe checks. I run the three calls against your AI agent personally, then send back a short findings note scored against the twelve points. Free: three calls cost me ten minutes, and if they surface a problem, putting it right is the work I charge for. Probing only ever happens with your written authorisation.

Include three things:

  1. Who you are, and the company.
  2. The number your AI agent answers on.
  3. One line confirming you authorise three test calls to it.

Reply within two working days. Voice first; if yours is chat-only, write anyway. linkedin.com/in/dylanevans

This page, and any findings note from the free probe, is a working checklist and a design review, not legal advice or a compliance assessment. The Commission's final guidelines on Article 50 were published on the 20th of July 2026; this page follows them. Checked against the Regulation and the guidelines on the 20th of July 2026.

Sources: Regulation (EU) 2024/1689 (Articles 50, 99 and 113) · Commission guidelines on Article 50 · Code of Practice on marking and labelling · AI Act implementation timeline

Who is checking: I have been casting and directing system voices since 2001, for SKY and the BBC, HM Revenue & Customs, the UK Census, and first-responder systems in Los Angeles. I have spent much of my career inside tightly regulated environments, including EU-regulation-driven programmes at Philips. An AI agent's opening line is a design decision. So is what it admits to being.