Every attendee,
answered.

Attendee questions are high-volume and time-critical, and the answers change during the event. A chatbot trained on last week’s PDF will confidently send people to a room that moved this morning. The useful version reads from the live systems: the agenda platform for sessions and rooms, registration for the attendee’s own ticket and bookings, and a venue guide the operations team keeps current.

So the concierge is retrieval over those sources, not a model’s memory. It answers what it can, says when it does not know, and hands anything else to the help desk with the conversation attached. During the event, a named person watches that queue.

Matchmaking is optional and opt-in. Attendees who choose it get suggestions of sessions and people based on what they told you, and can see why each suggestion was made. Nobody is suggested to others without having opted in.

We rehearse the event before it happens. The test set is last edition’s real questions plus a set of deliberately hard ones: the room that moved, the session that was cancelled, the attendee who mentions feeling unwell in a question about lunch. Correct routing of the urgent ones is the gate; accuracy on live-agenda questions is checked by changing the agenda during the rehearsal.

Attendees are told they are talking to an automated assistant, as the EU AI Act requires, and how to reach a person. Dietary and accessibility details are visible only to the staff who need them.