From transcript to action: how Squirrel decides what to do
A walk through the pipeline — capture, structure, extraction, resolution, execution — and the design decisions at each step that keep an autonomous agent from doing something stupid.
"Extract action items" sounds like one step. In practice it's five, and four of them are where the quality actually comes from.
1. Capture
Squirrel joins the call, records, and produces a diarized transcript — every line attributed to a speaker. Attribution matters more than people expect: "I'll take that" is worthless without knowing who said it.
2. Structure
The raw transcript gets segmented into topics before anything is extracted. A 50-minute call is rarely about one thing, and an agent reasoning over the whole blob at once will confidently merge two unrelated threads into one wrong task.
3. Extraction
Within each segment we pull out the things that have consequences, and label them by kind:
- Decision
- Something was settled. It needs to be recorded and findable later.
- Action item
- Someone owes someone else work. It needs an owner, a destination, and ideally a date.
- Blocker
- Work that can't start until something else moves. Worth surfacing, rarely worth filing.
- Open question
- Explicitly unresolved. The most commonly lost category, and the most expensive.
4. Resolution
This is the hard part, and it's what separates a note-taker from an agent. An extracted item says "Priya will update the onboarding flow before the demo." To act on that, Squirrel has to resolve four things:
- 1Who — which Priya, and what's her identity in the destination tool?
- 2Where — is this a Jira ticket in the product project, a Trello card, or just a Slack nudge?
- 3When — "before the demo" is a real deadline if the agent knows when the demo is.
- 4Whether — was this an actual commitment, or someone thinking out loud?
Each resolution carries a confidence score. High confidence flows through. Anything ambiguous is presented as a suggestion for a human to confirm — one tap, not a form.
5. Execution
Resolved actions are dispatched to connectors: a Jira issue with the right project and assignee, a Slack message in the channel that owns the work, a row in a tracking sheet, a CRM note against the right deal. Each write is idempotent — re-processing a meeting never doubles your backlog.
Every action has a receipt
Nothing Squirrel creates is anonymous. Each item links back to the exact moment in the transcript that produced it, so when someone asks "where did this ticket come from?", the answer is thirty seconds of audio, not a shrug.
What we deliberately don't do
- No silent guessing. If the owner is ambiguous, we ask instead of assigning to whoever spoke last.
- No action on side conversations. Brainstorms are captured as context, not converted into a backlog.
- No irreversible writes without scope. Connectors are granted narrow permissions, and destructive operations are out of scope by design.
An agent that files ten tickets and gets two wrong is worse than no agent at all, because now someone has to audit all ten. Precision beats coverage until trust is earned.
The short version
- Extraction is cheap; resolution — who, where, by when — is the hard part.
- Every action carries a citation back to the moment in the transcript that produced it.
- Low-confidence items become suggestions, not actions.
Give your company a second brain.
Squirrel captures every meeting, remembers every decision, and files the follow-up work for you.