Action Items & Attendees
How action items, decisions, key topics, and CRM attendees are extracted from every transcript — and how to review and edit them.
What Gets Extracted Automatically
Every processed meeting produces four structured outputs beyond the raw transcript. Action items — tasks someone committed to, with an assignee and a deadline where those can be inferred from the wording. Decisions — things the group explicitly agreed on ("we decided to…", "agreed to…", "decision: …"). Key topics — phrases that came up frequently enough to matter, derived from capitalized multi-word mentions, quoted phrases, and heading-style lines. And a short summary — the first substantive sentences of the meeting, capped at around 500 characters. The extractor uses pattern-matching heuristics (I'll…, we need to…, let's…, action item: …, please…, can you…, assigned to X: …) rather than a separate AI call, so extraction is fast and deterministic.
Assignees and Deadlines
The extractor tries to infer who owns each action item and when it is due. Assignees come from two places: explicit "assigned to X:" patterns, and implicit patterns where a capitalized name precedes "will", "should", or "is going to" ("Mark will send the SOW"). Deadlines are pulled from phrases like "by Friday", "before end of week", "due April 25", or short date forms. When the extractor cannot confidently infer either field it leaves them blank rather than guessing. Your AI employee can fill in the blanks for you if you ask — "assign the SOW action item to me and make it due Friday" — and the update is persisted.
Reviewing extracted action items
Look at what came out of a meeting.
Decisions and Key Topics
Decisions get their own field because they are strategically different from action items — they are commitments about direction, not tasks to complete. Capturing them separately makes it easy to look back at "what did we decide about pricing in the March meetings?" without wading through the action-item noise. Key topics are the conversational centers of gravity in the transcript, useful for building a high-level picture of what the meeting was about. Both decisions and key topics are text-only — there is no separate tracking lifecycle (decisions do not have a "status"; they are historical records of what was agreed).
Attendees and CRM Linking
When you provide CRM contact IDs at upload time, each contact is stored as an attendee of the meeting AND each contact gets a "Meeting" activity logged on their CRM record. The activity summary includes the meeting title, duration, and short summary, so the meeting shows up in that contact's timeline alongside emails, calls, and notes. You can also filter your meetings list by contact ID — "show me every meeting I've had with Jennifer" — to see the full interaction history in one place.
If you forgot to link attendees at upload time, you can add them later by asking your AI employee to "link this meeting to contacts X, Y, Z". The CRM activities will be backfilled for each added attendee.
Editing the Extracted Content
The extraction heuristics will miss things and occasionally over-extract. That is expected, and easy to fix. Ask your AI employee to "add an action item: rewrite the onboarding email by Monday, assigned to me" and it will insert a new action item into the meeting record. Ask it to "remove the action item about scheduling the standup — we agreed to handle it async instead" and it will delete the entry. Ask it to "change the assignee on the SOW action item from me to Sarah" and the assignee is updated. Every edit is just a sentence.