Observer listens to your meeting, holds the context, and speaks only when asked: the question nobody put, the number that doesn't add up, the option that breaks a constraint you agreed an hour ago.
Not a recorder. Audio is deleted as it is transcribed. Runs on your own server with your own model accounts.
The Atlas renewal will probably slip into Q4. Their legal team hasn't signed.
With Atlas in, we land the quarter at ninety-four per cent of plan.
Which gives us room to lift the hiring freeze in November.
Illustration of a staged review meeting. Names and figures are examples.
Most meeting AI writes notes after the fact. Observer works during the meeting. Tell it what the meeting is for, and it reads the room against that.
A title, and what you want from the meeting: settle a disagreement, check the numbers, widen the discussion. One line is enough. Names and roles of the people in the room and a few background figures are optional and make the answers sharper.
In the room through the laptop's microphone, or in an online call through the shared meeting tab (Meet, Teams, Zoom). The meeting owner announces it. It transcribes as the meeting goes, in pieces of under a minute, and keeps no recording. It keeps listening while you present slides.
Type a question when the room goes round in circles, or let it raise a quiet hand at an interval you set. It answers with grounds quoted from the transcript, with timestamps, or says it has nothing to add. One press reads the answer aloud.
These are the one-tap starting points inside Observer. Pick one, or write your own in a line.
Lay out the options, what they are weighed on, who isn't on board, what is still open.
Notice where people talk past each other, name what each side needs, suggest a way forward without taking sides.
Check what is said against the numbers and goals: slippage, risks, anything glossed over.
Bring in angles nobody raised, connect ideas across people, question assumptions.
Track what each side wants, what was offered or conceded, what is unanswered.
Say what you want from the meeting in one line.
“Settle the H2 pricing decision before we leave the room.”
Typical meetings: a quarterly business review, an H2 pricing decision, a hiring-plan debate, a vendor negotiation, a planning offsite. Made for leadership teams of 5–15 people whose meetings decide things; the person running the meeting starts Observer, and everyone else just talks.
Not built for: one-to-ones, large all-hands, or meetings where nobody can announce that an AI is present.
Eight features, each one shipped and in use.
Anchor, facts, assumptions, constraints, options, agreed, open. Inferred as the meeting runs, so it can spot an option that breaks a constraint or an assumption being treated as fact.
A short summary every ten minutes, so a two-hour meeting still has a shape and a question at the end can reach the start.
Each answer quotes the turns it rests on, with timestamps, and says what data it was missing. A blank is a blank, not filler.
On a timer, it checks the last stretch for something worth raising and shows one flag at a time. Ignore it and it goes away.
A firm, clear voice for when the room should hear it rather than read it. Off by default; one press per answer.
Speaker names from the transcription, corrected by you mid-meeting, and rechecked across the whole transcript at the end.
A laptop in the meeting room, or a shared tab in Meet, Teams or Zoom. It keeps listening while you present slides.
For a meeting that must not be kept: nothing is written to disk, the transcript lives in memory and is wiped at End.
Also in v1: personal accounts with admin and member roles; a Meetings page to reopen a summary, start the next meeting the same way, download the log or delete it; a per-meeting cost estimate; Hindi and mixed Hindi-English speech transcribed as spoken.
Audio reaches the server in pieces of under a minute, and each piece is deleted once transcribed. There is no playback.
It listens only between Start and End, and only in the tab it was started in.
Meetings are files on your own server. Only the person who ran a meeting and the install's admins can open it, and either can delete it, immediately and for good.
The transcript is wiped at End or after idle time. The AI services still receive the audio and text while the meeting runs.
“Quick note before we start: an AI Observer is listening to this meeting. It transcribes what is said and only speaks when one of us asks it a question. It is not recording audio; the audio is deleted as it is transcribed. If anyone wants it paused at any point, just say so.”
No. It transcribes in pieces of under a minute and deletes each piece. There is no playback.
No. It runs in a browser tab on a laptop in the room, or listens to a shared meeting tab online.
The person who ran it and the install's admins. Either can delete it.
The customer chooses: Google, ElevenLabs or Sarvam for transcription; Anthropic Claude or Fireworks for reasoning; ElevenLabs or Sarvam for the voice. On the customer's own accounts.
Hindi and mixed Hindi-English speech is transcribed as spoken.
Yes, every time. The product gives you the one line to say.
Pilot terms are agreed per team. Inside the product, each meeting shows its AI model cost; a typical hour is well under a dollar of model usage, billed to your own accounts.
Observer is v1 and in use for real leadership meetings. It is available as a pilot, installed on your own server.
Tell us the meeting you have in mind, who runs it, and where it happens. We will set up a pilot on your server, with your model accounts.