Does the assistant have to be labelled as AI?

Yes. Anyone interacting with an AI system must be able to recognise it — the EU AI Act requires this. The assistant therefore states visibly in the chat window that an AI is answering, and offers handover to a human at any time.

What the transparency duties mean in detail and what proper labelling looks like is described in the article EU AI Act: transparency duties for AI chats.

The label does not stand alone. Sourced answers name the passages they rely on, and the handover to your team is a separate step in the transcript. On sensitive topics the assistant holds back and offers the handover instead of formulating information; such conversations appear in the evaluation as escalations, separate from the self-resolved ones.

If your assistant requires consent, the snippet on your page carries the corresponding attribute; without it, the setting does not reach the host page. If consent is absent, no row is created in the database, because the check happens before the first write and not only on submission. In the window the visitor can request information, withdraw consent and delete the conversation; the widget does not bring its own consent banner. The basis for this is on privacy and hosting.