Can every meeting turn itself into decisions and tasks?

Hypothesis

Meetings produce transcripts; transcripts produce nothing. If an agent reads every new transcript and extracts the summary, decisions, and tasks into the systems where work actually happens, the meeting's output should survive the meeting.

What we built

A pipeline that watches for new meeting transcripts (from an AI notetaker or dropped files), analyses each one — executive summary, key topics, decisions, action items — then writes the analysis back onto the meeting page and syncs decisions and tasks into their own databases with proper status. Processed meetings are indexed so the analyser never double-handles and can improve its own extraction over time.

Learnings

• Decisions are the underrated artifact. Everyone builds task extractors; a searchable log of what was decided and when settles arguments months later.
• Idempotency is the difference between a pipeline and a mess — every stage needs a watermark so re-runs are safe.
• Extraction quality compounds when the analyser keeps an index of its own past outputs and corrections.

Log

• 2026-07-05 — Decisions + tasks syncing into their databases with status handling.
• 2026-06-20 — First transcripts analysed end-to-end.