Solution · reliability and meetings
Zoom calls → decisions, tasks and analytics: do your meetings turn into work?
Decisions sink into call recordings, the same questions come up again and again, and nobody knows which agreements became tasks. This solution picks up every meeting from Zoom, AI extracts decisions, commitments, blockers and open questions with a link to the exact place in the transcript, matches the decisions against tasks, and every week shows the lead what is stuck.
What it looks like
Mock-ups with demo data, no real data
01
Zoom
Recording and transcript
02
AI analysis
Decisions, commitments, blockers
03
Task matching
Did it turn into work?
04
Indicators and reports
Week, month, quarter
🎙️
Weekly sync, 24 September — analysis
- Participants
- EthanMariaAnna
- Duration
- 48 min
Decisions
- Move the newsletter launch to ThursdayTask found
- Add a second pricing tierNo trace
Blockers
- No access to the payments dashboard3rd mention
Open questions
- Who leads the new group?[line 214]
Meetings · week
Week of 22–28 September
Meetings: 14 · hours: 11.5
Decisions → tasks: 23 of 31
Needs attention
· Blocker “dashboard access” — 3 weeks
· 4 commitments overdue
· 2 questions older than 21 days
08:00
At a glance
- In production
- At a client under NDA
- Fits
- Teams that are on calls a lot and lose track of agreements
- Needs
- Zoom with cloud recording and Notion; tasks in Notion for matching
Estimate for your company
≈5 h
a week of meeting notes and task hand-out
Assumption: 20 calls a week × ~15 minutes of manual notes (our assumption)
This is a calculation, not a promise: plug in your own volumes and the figure changes.
Before
- Agreements stay in a recording nobody replays, and a week later they are retold from memory.
- The same decision gets made twice, and a blocker comes up at every meeting because nobody closed it.
- It is unclear which decisions became tasks, which have no owner and which promises are overdue.
- The lead cannot see how much time the team spends in meetings and whether they produce anything.
What was built
Automatic call capture
The recording, a transcript marked at each change of speaker, a summary and a Google Doc land on the meeting’s Notion card, linked to the calendar event and participants. Re-runs create no duplicates.
AI analysis of every meeting
Decisions, commitments with deadlines, open questions, blockers of seven types — each with its speaker and a link to the transcript paragraph. The rules separate “discussed” from “decided” and fact from hypothesis.
Matching decisions to tasks
Each decision is looked up in the task tracker; the final call belongs to an AI judge told that a false match is worse than a miss. Statuses: done, task found, needs review, no trace.
Memory across meetings
Decisions made again, recurring blockers and ageing questions build up a mention count instead of creating new records.
About 25 indicators
For the week, month and quarter: share of decisions that reached a task, decisions without an owner, overdue commitments, questions older than 21 days, meeting hours, meetings with no decisions.
Reports
On Monday, a Telegram digest with clickable “needs attention” items. At the end of each month and quarter, an AI review that is not allowed to invent numbers.
Reliability
A daily token budget, sampled re-checks on a stronger model, and “no data” kept distinct from zero.
📊Meeting indicators
| Indicator | Value | Signal |
|---|---|---|
| Decisions → tasks | 23 of 31 | 74% |
| Decisions without an owner | 3 | watch |
| Decisions made again | 2 | repeat |
| Questions older than 21 days | 2 | stuck |
| Meetings with no decisions | 4 of 14 | 29% |
Stack
- Zoom API
- n8n
- Notion
- Google Docs
- Telegram
- AI models via API
Where it runs
Questions
Is this just a meeting summary?
No. The summary is only for navigation. The core is structured decisions and commitments linked to the transcript, their matching against tasks, and indicators for the period.
How does the system know a decision became a task?
It searches for similar tasks by keywords and project, and the final call belongs to an AI judge with strict rules. Doubtful cases are marked “needs review” rather than counted.
Could the AI invent decisions that were never made?
Every item is tied to a transcript line and a speaker, the prompt forbids filling gaps, and a sample of meetings is re-checked on a stronger model.
What does the lead see?
Every Monday, a short digest: meetings and hours, how many decisions reached tasks, what repeats and what is overdue, with links into Notion.