Article 50 puts a sentence at the front of every call
The EU AI Act obliges you to tell callers they are speaking to a machine. That obligation lands in one specific place, the greeting, and it changes how the first three seconds of a call have to be built.
Article 50 of the EU AI Act is short, and for anyone running a voice agent it is the part of the regulation with the most immediate consequences. Systems that interact directly with people have to be built so that the person is informed they are interacting with an AI, unless that is obvious from the context. On a phone call it is not obvious. Sounding human is the product.
What makes this awkward is not the disclosure itself. It is that the obligation attaches to a moment rather than to a document. You cannot satisfy it in a privacy policy, in a recorded notice halfway through, or by answering honestly when asked. It has to be in the greeting, before the conversation starts.
Why the greeting is the hard case
Most voice platforms let the first message be generated by the model. That is usually the right call: a greeting that adapts to the caller, the time of day, or the reason for the call is better than a fixed line. But a generated greeting is a probabilistic greeting, and a legally load-bearing sentence cannot be probabilistic. The model will produce the disclosure almost every time, and almost every time is the wrong standard for a compliance control.
The second problem is interruption. Barge-in is a feature. A caller who starts talking over the agent should be heard immediately, and every decent voice stack cancels the outgoing audio the moment it detects speech. Applied to the disclosure, that feature means an impatient caller never hears it. The regulation does not have an exception for callers in a hurry.
The disclosure has to be deterministic text, prepended ahead of anything the model produces, spoken in a window where barge-in is suppressed.
How we implemented it
We treated it as a property of the call rather than a property of the flow, because a setting that an operator can forget to switch on is not a control. Accounts with their processing region set to the EU get the disclosure whether or not the flow asks for it, and the flow editor shows the toggle as forced rather than pretending it is optional.
- 01The disclosure resolves to fixed text from a catalog, one entry per supported language, rather than being generated.
- 02If the flow’s first message is plain text, the disclosure is prepended to it. If the first message was set to be model-generated, the greeting is replaced by the disclosure alone and the model picks up from the second turn.
- 03The greeting window is marked uninterruptible, so barge-in is suppressed until the sentence has finished.
- 04Mid-call state changes that re-enter the same flow deliberately skip it. The caller has already been told, and repeating it would be noise rather than compliance.
Fixing the third point turned up a bug that predated the feature. The flow editor had labelled plain-text greetings uninterruptible for a long time, and the local voice pipeline had never actually honoured the flag. Callers could talk over a greeting the interface promised they could not. That is now fixed for every flow rather than only for the ones subject to Article 50, which is the sort of thing that tends to fall out of taking a compliance requirement seriously.
On the wording itself
There is a real tension here. A catalog of pre-written phrasings is safer, because the sentence that carries the legal weight is not being improvised by whoever configured the agent that afternoon. But operators have counsel, counsel has opinions, and a platform that refuses to let a company use its own lawyer’s wording is not usable in the enterprise.
We landed on both. The catalog is the default and covers every supported language. You can also prepend or append to a catalog entry, or replace it entirely with your own text. Choosing your own wording raises a standing warning in the editor, because at that point the sentence is yours and so is the exposure.
Our catalog phrasings are drafted from public guidance, not from legal advice. Have your own counsel read the line you ship.
What this costs you
About two seconds at the top of the call, and a small number of callers who find it strange. In our experience the second cost is smaller than teams expect. People are increasingly good at spotting a machine, and being told up front reads as confidence rather than as a warning. The calls that go badly are the ones where the caller works it out on their own, halfway through, and feels they were being handled.