Frequently asked questions about AI avatars at events

It can happen. The moderator reads the wrong line on the tablet before the room hears it. One tap on STOP silences it in under half a second, even mid-sentence. The moderator can then correct the record or hand the floor to someone else.

It is set up not to. When it lacks a fact, it says so.

It stays silent until the moderator calls it by name or hands it a question on the tablet. Talking about it rather than to it triggers nothing.

Right after its answer, usually not. A plain “Und warum?” (And why?) is then enough, while a bare “Okay” gets no reply. The conversation ends 12 s after the last sentence, a value we set, or as soon as the moderator turns to someone else; after that, its name is needed again.

Yes. It has the last few minutes word for word. Everything earlier it summarizes as it goes, so far measured only in the lab. It addresses people by name.

Not its own. Its position comes from the briefing, such as the product’s or that of its critics. It pushes back, and when someone has a point, it grants it.

The moderator, through the wording. “Kurz” (briefly) or “ja oder nein” (yes or no) gets one sentence, “erklär uns” (explain to us) a thorough explanation of a good half minute, and anything else three to four sentences.

An interjection without its name goes right past it. If the moderator calls its name into its answer, it says a short German line such as “Moment — …” (one moment) and resumes at the start of the interrupted clause; that is measured in the lab, not yet live. If a guest should take over at once, the moderator taps STOP.

If it is told. Before the event, the names of the moderator and the guests go into the briefing on the tablet. In our live tests, it used them to address the guests correctly.

Albert provokes, Vera checks the facts, Tina explains, and Ken speaks for those not on the panel. For a panel we recommend Albert, Vera or Ken, at a university Tina, plus Albert or Vera for a contested question.

The topic, who sits on the panel and in what capacity, and the position the AI avatar should argue. Face, voice and attitude come with the character. If someone on stage has a name close to its own, it gets a different one when the character is chosen, so the new name is already in the invitation.

No. Its one language for now is standard German, which rules out English as well. Whatever is said on stage in Swiss German, it cannot pick up later.

After 1.5–2 s, measured live from the last word of the question to the first sound; a follow-up without its name takes a little longer. It opens with a German “Also.”, “Hm.” or “Nun.” (so, hm, well) and picks up from there in its first sentence. The floor is about 1.5 s, so we promise nothing under a second.

We do. One of us sits with the laptop beside your sound engineer at the desk, and the moderator gets the tablet. From your technicians we need the moderator’s microphone as its own channel, the other stage microphones via an aux send, a return into the desk for its voice and a screen for its picture.

About two weeks ahead. First we settle the character, the labeling and the date, and most of the time after that goes into working things out with your event technicians.

Then it goes silent, because speech recognition, the language model and the video all come from the cloud. The moderator moves on with a line rehearsed in advance.

In the EU, generally yes, since August 2, 2026, to protect the guests who talk to it; for the room, a notice is good practice. At an event in Switzerland, Article 50 is generally not binding. We recommend the notice anyway and can label the video on screen as an AI avatar.

Under the AI Act, it is one if it resembles an existing person and falsely appears authentic. Our four characters are computer-generated and modeled on no real person, so in our reading they are not deepfakes.

Not by default. It hears what your sound team sends from the desk, which is the stage microphones. Voices from the room reach the transcript only if your sound team also sends it an audience microphone, for instance for questions from the floor.

As a transcript on our laptop. On the way, the audio passes through speech recognition in the cloud, and to write its answers, the text goes to a language model, also in the cloud.

Because our offering is in beta. The technology is built and measured live, in tests on a laptop with people in the room speaking freely. If you plan an event with us now, it runs as a pilot.

We have no public price list. We offer the first engagements on pilot terms. Send us the date and type of your event, and we will tell you the terms.