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.

Scale
Small
Platforms
Web · Slack/Teams surface
Capabilities
AI · Internal tools · B2B

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.

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?

Facing a similar problem for real?

This study's reasoning — discovery, architecture, evaluation — is exactly what a HummingByte engagement looks like. Bring us the real version.