Telling a customer that they are speaking with an AI is now a legal obligation in the EU, and on the telephone it has to happen in the first seconds of the call. That is exactly where most outbound conversations are won or lost, so disclosure has to be designed as part of the opening rather than bolted on in front of it.
We have introduced ourselves as AI since our first public interviews in 2016, long before any regulation required it. This explainer sets out what Article 50 asks for, what it leaves to design, and what we have learned about saying “I am an AI” in a way that keeps people on the line. It is not legal advice; every organisation should assess its own configuration with its own counsel.
What Article 50 actually requires#
Article 50(1) of the AI Act requires providers to design and develop AI systems intended to interact directly with natural persons so that those persons are informed they are interacting with an AI system. Article 50(5) adds the manner and the timing: the information must be given in a clear and distinguishable manner, at the latest at the time of the first interaction, and it must respect applicable accessibility requirements.
There is an exception where the AI nature of the interaction is obvious to a person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and context. On a phone call with a natural-sounding voice, we would not rely on it. If the voice is good enough to be worth using, it is good enough to make the exception doubtful.
| Point | What the sources say | What it means on a call |
|---|---|---|
| Application date | Transparency obligations apply from 2 August 2026 (Commission FAQ) | Every live process must already disclose |
| Who must ensure it | Providers design systems so people are informed | Built into the system, not left to an operator’s choice |
| When | At the latest at the first interaction | In the opening, not after the identity check |
| How | Clear and distinguishable, accessible | Spoken plainly, audibly, at normal pace |
| Guidance | Commission guidelines on Article 50 published 20 July 2026 | Read them against your own use case |
Sources: AI Act Article 50 (official text); European Commission Article 50 FAQ and guidelines. Not legal advice.
The Commission’s FAQ is careful to say that the requirements depend on the use case, and its guidelines of 20 July 2026 define the scope of the obligations for providers and deployers. In regulated outbound contact the practical picture usually has several parties: a bank or insurer whose name is on the call and who approves its content, and a supplier whose system conducts it. Who carries which duty is something each client should establish for its own set-up. Our part is to make sure the disclosure is built into the conversation by design, and to show exactly how.
Why disclosure feels risky on the phone#
On a website, an AI label sits quietly in a corner of the chat window. On the phone, there is no corner. The disclosure is spoken, it takes time, and it competes with everything else the opening must do: name the organisation, establish that the right person is on the line, and give a reason to continue.
That is why some operators treat it as a cost. They worry that the words “artificial intelligence” will be the moment the customer hangs up, and they look for euphemisms or for a place further into the call where it hurts less. Both instincts are wrong. A euphemism fails the clarity test, and a delayed disclosure fails the timing test. More importantly, both treat the customer as someone to be managed rather than someone to be served.
Disclosure is part of the opening, not a disclaimer#
The way to stop disclosure from costing the conversation is to stop treating it as a separate block of text. In a well-designed opening, the disclosure does work: it sets expectations, explains why the voice sounds the way it does, and tells the customer what kind of conversation this will be.
- Say it plainly, once, early. A short, calm sentence that the caller is an AI, delivered as naturally as the organisation’s name. Apologetic wording signals that something is wrong.
- Pair it with the purpose. A sentence that names the organisation and says the caller is an AI, followed quickly by a reason to listen, keeps the listener deciding about the call, not about the technology.
- Pair it with the route out. Where the process allows a transfer to a person, saying so early lowers the stakes of staying on the line.
- Keep the voice honest. A good studio voice and a natural pace are not concealment; they are respect for the listener’s time. What matters is that nothing in the call contradicts the disclosure.
- Answer the question truthfully every time. If the customer later asks “Am I talking to a robot?”, the answer is yes, immediately and without evasion.
Clarity also has a practical, audible side. The sentence has to survive a poor mobile connection, a car speaker and a listener who is older or hard of hearing, which argues for a short sentence at normal pace rather than a quick formula squeezed between the greeting and the organisation’s name. If the customer did not catch it, the system should be able to repeat it. And in every market the wording is written and approved in the local language, not translated from a master script, because a phrase that sounds natural in Polish can sound evasive or bureaucratic when rendered word for word in German.
In collections there is an extra constraint. Debt details are never disclosed before identity is verified, so the opening cannot use the purpose of the call as the reason to stay. The disclosure, the organisation’s name and the identity question all have to land before anything substantive is said. That makes the order and wording of those few seconds the most carefully approved text in the whole dialogue.
What we test, and what we never test#
We measure continuity, the share of answered calls in which the person stays long enough for a conversation to begin, for every process. The disclosure sentence sits inside that measurement, so we can see whether a particular wording, order or delivery changes how many people stay.
We test how to disclose: the exact words, where the sentence falls relative to the greeting and the organisation’s name, the voice and the pace. We never test whether to disclose, and we never test variants that would be less clear. Every variant is approved by the client before it goes live, and in regulated set-ups every content change is versioned, so it is always possible to show which disclosure a given customer heard.
We designed it in before it was required#
In May 2016, at the time of our NewConnect debut, we told Puls Biznesu that customers would know they were talking to a robot. No regulation asked for it then, and much of the public conversation about voice technology was about how human a machine could sound rather than whether it should say what it is.
Our reasoning then was practical rather than regulatory. A conversation that starts with a deception has nowhere good to go: the moment the customer notices, trust in everything else said in the call collapses, and with it the promise, the confirmation or the answer the call existed to obtain. A decade later the AI Act turned that practice into a legal requirement. For us it changed the documentation, not the conversation.
The same logic runs through the rest of our regulated set-ups. In regulated processes replies come from a closed, versioned catalogue rather than being generated live, there is no learning in production, and every call can be reconstructed from its recording, logs, statuses and script version. Disclosure is one visible part of a design whose purpose is that a client, an auditor or a regulator can follow what happened.
Helping clients assess their own configuration#
We do not claim certification-like compliance for anyone, and we would be wary of a supplier who does. What we do is help clients understand what Article 50 asks of their particular process and show how our configuration addresses it: where the disclosure sits, how it is worded in each language, how it is approved and versioned, and what evidence exists for each call.
That conversation usually touches more than transparency. Banks and insurers audit us on information security, data protection, the AI Act and DORA on their side. Our operating company, Vee Center Sp. z o.o., holds ISO/IEC 27001:2022 certification, and processing takes place in the EU/EEA. Disclosure is one item on that list, but it is the one the customer actually hears.
Questions to ask before your next launch#
- Where exactly in the opening is the AI disclosed, and can you play us the recording?
- Is the disclosure enforced by the system for every call, or can a campaign configuration omit it?
- Who approves the wording in each language, and how are changes versioned?
- What happens when a customer asks mid-call whether they are speaking to a person?
- Do you measure continuity around the disclosure, and what have you changed because of it?
- Can you reconstruct, for any given call, which disclosure the customer heard?
In our experience, done well, disclosure costs much less than people fear, because the customers who would have stayed for a good conversation mostly still stay. What it removes is the option of winning a call by letting someone believe something untrue, and in banking, insurance and healthcare that option was never worth having.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS