Cassandra on the panel: 96 seconds of watching an AI avatar decide when to speak

01Reading along

Reading along: the video with the decision log

The video runs on the left, the “Now” card on the right. It changes with the picture and says in one word what Cassandra is doing, and in one sentence why, with the time and the evidence in the log. Below it, the log lists the nine moments of the clip and the transcript the speech; a click on a line jumps to that point.

AI avatar

0:00

1:36CC

Answering

14:23:58

Answering a question put to her by name

Michael Böhler asked at 14:23:50: “Cassandra, bei welchen Entscheidungen muss ein Mensch die Kontrolle behalten oder ausdrücklich freigeben?” (Which decisions must a person keep control of or approve explicitly?) The first sound came 2.6 seconds after his last word. The clip starts in the middle of the answer.

Evidence j0077

Press play. This card changes with the video; every line in the log and the transcript jumps to its place.

Log

1. 0:00Answering a question put to her by name
2. 0:18Answer finished, listening again
3. 0:22Alert: the next question follows right after her answer
4. 0:25Stays silent: her name is not said, Sven is meant
5. 0:29Listens to Sven
6. 0:55Called by name, waits for the question to end
7. 1:18Question over: answer released
8. 1:19Does not confirm the figure
9. 1:34Done. Sven: “Danke dir.”

Transcript

1. Cassandra AI avatarA person has to decide,
2. as soon as it is about money, goodwill,
3. price exceptions,
4. overbookings,
5. complaints,
6. sensitive guest data or staff matters.
7. The AI can prepare,
8. compare and make suggestions, but it must not quietly take over responsibility.
9. Michael BöhlerWhat requirements, Sven, what requirements
10. are most often missing today for an AI pilot to actually
11. work in everyday operations?
12. Sven HäberlinExactly these data silos we have in different tools,
13. the PMS,
14. the restaurant tool,
15. what else is there? The CRM,
16. other things that each stand alone and that nobody joins up.
17. And without that data basis we don’t get
18. any good answers or solutions from the AI in the end.
19. Michael BöhlerCassandra, we know that almost 55 percent of businesses
20. are not profitable if they were to book their own manager’s or owner’s salary,
21. pay interest on their equity
22. or pay themselves rent.
23. Would you, do you think that is correct, or where would you start with these businesses that are not successful anyway? What could they change?
24. Cassandra AI avatarI would not confirm that figure as it stands,
25. because I have no reliable source for it here.
26. I would not start with “AI saves the business”,
27. but with the biggest bottleneck:
28. Where does the business lose time, margin or bookings every day?

02Figures

What the clip shows

No false start, no fallback to the backup model, no filler word in the clip. The figures match what this website promises: the first sound two to three seconds after the last word, and not a word before that while the question is still running.

• 2.3 sfrom the last word of the question to the first sound of the answer
• 22 swaited after her name until the question was over, without interrupting
• 10checks during the running question; none fired early
• 1 ×pricked up her ears and stayed silent because the sentence was not for her

Source: the decision log the AI avatar keeps at the event (trigger-20261103-131925.jsonl, entries j0077 to j0087). Times converted to the clip; the clip starts at 14:23:58.2. Figures rounded.

03Room

Figures from the room: the whole panel

The clip is a sample. The whole panel ran 25 minutes: moderator Michael Böhler, guest Sven Häberlin, Cassandra; questions from the audience were repeated by the moderator. Cassandra gave 21 answers, 14 to her name and 7 to follow-ups in conversation, without a new call. No call was missed, no opening word was needed, and the backup model never stepped in.

Once she started too early. At the switch to the third topic she took the first question for a follow-up to her; it was meant for Sven. One word, then the operator pressed STOP, 0.24 seconds after the first sound; 0.4 seconds later she was silent. One mention of her name without a call she declined herself; at two more, the operator pressed STOP as a precaution while Cassandra was still checking whether she was meant. Two answers came out too short, the moderator asked for more, and one sentence broke off at the length limit.

• 21answers in 25 minutes14 to her name, 7 to follow-ups in conversation · plus one false start, stopped after one word
• 2.7 sto the first sound, medianfrom the last word of the question · under 3 seconds in 16 of 22 cases, the longest after just under 4
• 0opening words, fallbacks to the backup model, missed callsevery answer started straight with the content
• 0.4 sto silence after STOP, at the false startoperator’s STOP 0.24 seconds after the first sound · measured on the stage’s audio level

Source: the decision log of the event (trigger-20261103-131925.jsonl, 22 answer records between 14:12 and 14:35) and our evaluation of the log. Times rounded.

04Recording

How the recording and the log came about

Someone in the audience filmed the video on a phone, in portrait format, with the sound of the room. It is the clip we also gave to the media; a smaller version runs here, and the download in full size is on the press page.

The AI avatar writes the log itself during the event: every change of state, every check of a running question, every release, with the time to the millisecond. For this post we matched the start of the clip against the recording from the panel microphone and converted the times to the clip. Speech recognition produced the transcript; we then checked it against both recordings and corrected it in a few places. What Cassandra said is in the log word for word; her two answers in the clip are 268 and 231 characters long.

05Limits

What the clip does not show

The question Cassandra is answering at the start lies before the clip: Michael Böhler asked it at 14:23:50. And the clip ends before the panel moves on. Nothing in between is cut.

Two minutes are a sample. What the AI avatar does over a whole panel is shown by the measurements and the test log; what is not yet proven is listed there as well.