Turned two years of recorded meetings into an askable knowledge base for a 25-person consultancy
Client snapshot
A 25-person management consultancy. Partners run client workshops and steering meetings; associates write them up. Every meeting gets recorded, and the recordings land in a shared Drive folder as video files.
The problem
The recordings were the most complete record of every engagement, and nobody used them. Finding what a client said about their budget in a March workshop meant guessing which file it was in and scrubbing through an hour of video. Associates spent their first weeks on a project rewatching meetings. When a partner left, what they knew left with them.
What we built
A watched folder and a pipeline behind it.
When a new recording lands in the Drive folder, n8n picks it up, strips the audio with ffmpeg and sends it to Whisper for transcription. Claude reads the transcript and writes a one-page summary: who attended, decisions made, open questions, action items with owners. The transcript and summary are filed next to the recording under the client and project, with a consistent name.
Every transcript is split into passages and indexed in Postgres with pgvector. A Slack bot answers questions against that index: "what did the client say about phase two pricing?" comes back with a short answer and links to the meetings and timestamps it came from. If the bot can't find a solid answer, it says so and logs the question.
Access follows the engagement. A consultant only gets answers from meetings on projects they're staffed on.
Results, 60 days in
- Backlog of about 400 recordings transcribed, summarised and filed in the first week.
- New recordings filed with a summary within an hour of upload, with no one touching them.
- Finding a specific client statement: 20 minutes or more of scrubbing → under a minute.
- About 85% of questions answered with a source link; the rest landed in the weekly unanswered-questions log.
- Associates joining a project now read the summaries instead of rewatching meetings.
Quote
"New people used to lose their first week to recordings. Now they ask the bot and click the timestamp." Operations lead
Build snapshot
- Timeline: 16 business days
- Stack: n8n, ffmpeg, Whisper, Claude, Postgres + pgvector, Google Drive, Slack
- Investment: $6,500
- Who owned it: the operations lead owns the bot and reviews the unanswered-questions log each week.
The pattern
This is the pattern we ship in the Internal Knowledge Chat engagement. Here the source was meeting recordings; the same build works over Notion, Drive or Confluence.
This is an illustrative composite drawn from operator engagements. Names, identifying details, and exact metrics are anonymised. The workflow, stack, and approach are real.
Looks like your situation?
A 30-minute call tells us whether the same pattern fits your operation, and what it would cost to ship it.