Team productivity
AI Meeting Decision Tracker
Extract decisions and owners from meetings — not summaries — into a log that survives. A missed decision beats an invented one.
The experiment
Meeting summaries are the most generated and least read artifact in modern work. The durable output of a meeting is smaller and sharper: what was decided, and who owns it. The experiment asks whether extracting exactly that — typed decision records, not prose — produces a log a team still consults six weeks later, when “didn’t we decide this already?” comes up for the third time.
Precision over recall
The stance that shapes the whole build: a missed decision is a small, recoverable failure — someone adds it by hand. An invented decision is a poisonous one; a log containing even a few fabricated entries stops being a log and becomes another thing to double-check. So extraction runs at a deliberately high confidence threshold, anything ambiguous lands in a short “possible decisions” queue for one-click confirmation by an attendee, and every confirmed record links to the transcript moment it came from. The log’s value is that it can be trusted blind; recall is negotiable, precision is not.
Consent before capture
Recording ethics are a launch requirement, not a settings page. Capture is per-meeting opt-in by the organizer, announced visibly to every participant, with all-party consent required where the team’s jurisdiction demands it. Transcripts are means, not product: they are retained briefly for the confirmation flow, then deleted — the decision records, with their short source excerpts, are what persists. No ambient capture, ever.
Shape of the build
- Ingest: meeting transcripts from the team’s existing conferencing stack
- Extraction to a typed record: decision, owner, date, meeting, source excerpt
- Confirmation queue surfaced in Slack/Teams; edits and additions welcome by design
- The log itself: searchable, filterable by owner and topic, linkable from anywhere
Evaluation
- Extraction precision on a labeled transcript corpus — the gating metric, with a zero-tolerance audit for records that have no basis in the source
- Confirmation friction: seconds per queued item; if confirmation feels like homework, the queue will rot
- The survival metric, which decides the experiment: do teams consult the log unprompted after week four — lookups, links pasted into chat — or does it join the graveyard of meeting artifacts?