Context
The client was a fast-scaling product organisation that had grown from 30 to 80 people in fourteen months. With engineering in Europe, design in North America, and leadership in Asia, the team had built a healthy meeting culture and an unhealthy meeting load. The CEO described it bluntly: ‘We are making good decisions and then losing them.’ Recordings were piling up in shared drives, async folks were learning about closed decisions days later, and the company was pouring energy into work that had no memory.
The Problem
- Individual contributors were losing roughly 14 hours a week to meetings and meeting-adjacent work (prep, follow-up, recap reading).
- Action items were disappearing into Slack threads with no owner, no due date, and no aggregated view.
- Asynchronous team members were consistently joining conversations after decisions had already been ‘closed’ — hurting morale and slowing execution.
- Meeting recordings were piling up in cloud drives that almost no one re-watched.
- Leadership had no reliable way to ask ‘what did we decide on pricing in Q3’ and get an answer in less than an hour.
Our Approach
We started by ranking meeting types by business value: weekly product reviews, design critiques, all-hands, and customer calls. The agent was rolled out in that order, with each tier validated before the next was enabled. Every output — brief, action item, Q&A response — was designed to be skimmable in under 60 seconds, because nobody was going to read another wall of text.
What We Engineered
- Real-time transcription of every Zoom, Meet, and Teams call the team opted in to, with speaker labels and timestamps preserved.
- Structured briefs that separate decisions, owners, due dates, open questions, and risks — not a wall of summary text.
- A retrieval-augmented archive so anyone in the company can ask ‘what did we decide on pricing in Q3’ and get a cited answer with the meeting clip attached.
- Auto-creation of Linear and Jira tickets for every action item, tagged to the right owner and posted into the relevant channel.
- A weekly ‘what changed’ brief delivered to leadership every Monday morning.
Architecture and Workflow
- Interface: Slack bot for daily use plus a lightweight web dashboard for search and review.
- Models: GPT-4o for summarisation and action-item extraction; Whisper for high-accuracy speech-to-text.
- Inputs: Zoom, Google Meet, and Microsoft Teams recordings; uploaded transcripts; Slack thread exports.
- Stack: FastAPI, pgvector for the searchable archive, Linear and Jira webhooks for ticket creation, scheduled daily and weekly brief jobs.
- Outputs: Markdown briefs, structured tickets, cited Q&A responses, weekly leadership digest.
- Time to launch: Easy — two weeks end-to-end.
Business Impact
- Async engagement on key product decisions increased by 73% — measurable in Slack threads and document comments.
- Action-item completion rate moved from 41% to 92%, with every item tied to an owner and a date.
- Time spent re-watching meetings dropped by 88%.
- Recovered approximately 2.6 productive hours per individual contributor per week.
- First-time leadership decisions about hiring, roadmap, and pricing were searchable, cited, and discoverable within minutes.
Why It Worked
Meeting tools are everywhere. Meeting intelligence is rare. We did not try to replace the way the team met. We engineered an agent that turned the noise of meetings into structured, accountable, searchable output and we shipped it before the team's enthusiasm cooled.
Risks Mitigated
- Privacy — opt-in per meeting, with clear retention windows and PII redaction at storage.
- Tool fatigue — delivered inside Slack, not as another tab.
- Hallucination — every claim in the archive is cited to a transcript timestamp.
Client Voice
“We did not buy a meeting tool. We bought our team's time back. The Monday brief is the most-read message in the company now.”
— Chief of Staff | Distributed Product Organisation