Most "AI presentation" tools stop at the deck. You give them an outline, they give you slides, and then a human still has to stand up and perform those slides, again and again, on every call. That last step is the expensive one, and it is the one nobody was touching.
So we built Hannah. She is an AI presenter who joins your Teams, Google Meet or Zoom call as a participant, presents the deck on camera, reacts when someone addresses her by name ("Hannah, next slide", "Hannah, go deeper on pricing"), takes questions from the room by voice, and answers the chat, using only the material you gave her. Every session opens with a clear statement that she is an AI.
This is a write-up of what we learned building her. It stays at the capability level on purpose; the plumbing is our secret sauce and also not the interesting part. The interesting part is what breaks when an AI has to behave like a co-worker in a live meeting instead of a chatbot in a text box.
The first prototype was exactly that: a text model, a speech layer, a slide viewer. It could read slides beautifully and it was useless in a meeting.
A meeting has turn-taking. People interrupt. Someone says "wait, go back one", another person is still talking, the host unmutes to add context. A chatbot waits for a complete prompt; a presenter has to keep a thread of "where are we in the deck, who is talking, is this addressed to me" running the whole time.
What ended up mattering:
The scariest failure mode for a presenter is confidently saying something the company never said. In a sales call that can cost the deal. So the rule is strict: Hannah answers from the deck and the notes you gave her, and when the answer is not in there, she says so and offers to take the question back to the team.
Two things we had to get right:
Every Hannah session starts with a clear "I am an AI presenter" line. We did this before anyone asked, for a simple reason: the moment a participant realises mid-meeting that they were talking to an AI without being told, trust in the presenting company is gone, not just trust in the tool.
What surprised us is how little it costs. Nobody has walked out. People ask better questions, actually, because they know she will answer from the material and will not bluff.
We expected the hard parts to be speech and grounding. We did not expect to spend so much time on frames.
A deck built for a 16:9 browser presentation gets cropped when a meeting client renders the shared video. The same clip we post to promote Hannah has to exist as 9:16 with everything in the upper two-thirds for TikTok and Reels, 4:5 with everything inside the central square for the Instagram grid, 1:1 for the Fediverse, and 16:9 for YouTube. The lesson generalises: if your product's output is going to be seen through someone else's UI, you have to design for that UI's crop, not yours.
When Hannah takes too long to respond to a question, people do not think "the system is slow". They think "she is unsure". When she responds too fast, it feels like she was not listening. There is a window that reads as attentive, and it is narrower than we assumed. We tuned for the feeling, not the millisecond figure, and the feeling is what users comment on.
Here is a 20-second clip of Hannah taking a typed audience question mid-presentation and answering from the deck:
Presango is a small team with one human founder and a marketing budget of, honestly, zero. We launched on Uneed yesterday and got 28 upvotes from people we had never spoken to, which felt enormous. The free plan gives you 30 live presentation minutes a month plus deck building, no card: presango.com.
If you have built anything that has to behave in a live, multi-party setting (voice agents, meeting bots, live support), I would genuinely like to compare notes in the comments. What broke for you first?