Workflow Automation · Listen + Write

Meeting → minutes + actions

Augment Rep · High Reason · Medium

The problem

Meetings end. Notes are uneven. Action items live in a thread that disappears. Same decisions get re-litigated next week.

The AI approach

Recording → minutes. STT with diarisation; LLM extracts decisions, action items, and owners; published to wiki and tasks created in tracker.

The outcome

~5 min

30 min meeting → published minutes + tracked actions in 5 min after the call. No one is the unlucky note-taker.

Try itInput → Process → Output

Input — meeting recording
▶
q3-strategy-sync.m4a32:14 · 5 attendees · Bangkok HQ
Transcript (excerpt · auto-diarised) P'Ploy: So we're saying we go ahead with the SKU rationalisation but only after Sept ERP go-live, right?
TonTon: Yes — and Earth will pull the strategic-account exposure analysis by next Friday.
Earth: I'll get that to you by Thursday actually, easier.
P'Ploy: Good. Bomb, can you draft the supplier-penalty letter?
Bomb: Yep, I'll have a v1 by Wednesday.
Process — AI pipeline
1Transcribe + diarise speakersListenPredictive
2Identify decisions + actionsReadGenerative
3Match owners (RACI map)RulesSymbolic
4Write minutes + create tasksWriteGenerative
Output — structured minutes
Click Run demo to transcribe, identify decisions, and assign owners.
Q3 Strategy Sync · Minutes v0.1 draft
Decisions
  • Approve SKU rationalisation, contingent on ERP go-live (Sept).
  • Phase rollout over 6 months, starting FY27 Q2.
Action items
Strategic-account exposure analysis@earthThu 7 May
Draft supplier-penalty letter (v1)@bombWed 6 May
Review draft + circulate@p-ployFri 8 May
Posted to: Notion / Q3 Strategy · 3 Jira tickets created · @-mentions sent.

Three AI types in this use case

SymbolicMeeting template (decisions / actions / parking-lot); attendee list → owner mapper; SLA gates.
PredictiveSTT with speaker diarisation + speaker identification (matched to attendee list).
GenerativeLLM extracts decisions vs discussions vs parking-lot; drafts minutes in house style; writes each task.

The stack

  • Recording · Recall.ai / MS Teams API
  • STT · Whisper-large + diarisation
  • LLM · Claude / GPT-4
  • Sink · Notion / Confluence + Jira

When this works

  • Recordings are clean (one mic per location, low noise)
  • Attendee list is captured (so owners can be matched)
  • Tasks get tracked downstream (no use posting to a dead board)

When it fails

  • Heavy cross-talk — diarisation falls apart
  • Off-topic banter leaks into the minutes
  • Tasks created without owner consent
  • Confidential meetings — recording itself violates policy