"Someone should do that" is not a task
Meetings generate commitments—"Legal will review," "Sam to update the estimate," "Send client the revised timeline." Without capture, they live as fading memory and Slack pins. Sprint boards miss them; decision_inbox never sees them; next week the same topic returns.
AI extract meeting action items accelerates capture—but execution requires tasks in the system of record. WKFGo create_task via MCP (or UI) creates assigned work with project context, due dates, and audit trail after humans confirm each item.
Framework: from utterance to owned task
1. Extract candidates (AI)
From transcript or notes, AI lists:
- Verb + deliverable ("Draft rollback plan")
- Named owner if spoken
- Implied deadline if mentioned
- Related project if context clear
Mark confidence: explicit vs inferred—inferred items need facilitator confirmation.
2. Human review (required)
Facilitator checks:
- Correct owner—don't assign to someone who was only mentioned
- Right project and sprint boundary
- Not duplicating existing work—
smart_searchfirst - Sensitive items routed correctly—not public task titles
3. create_task (system of record)
For each approved item, create_task with:
- Title humans would say aloud
- Assignee
- Due date (even rough—better than none)
- Link to meeting wiki page if
publish_meeting_wikiran
4. Close the loop
Next standup: list_tasks filtered by "created since steering" or tag. Unfinished items escalate—not re-debated from scratch.
Pair with meeting summarization
Best practice sequence:
- Meeting ends
- AI drafts summary + action list
publish_meeting_wikifor decisions narrativecreate_taskper confirmed action- Comment on tasks linking wiki slug for context
Decisions live in wiki; work lives on the board—both searchable via smart_search.
AI prompts for extraction
Via WKFGo MCP:
- "From these notes, list action items with owner and due date—flag inferred owners."
- "smart_search for 'rollback plan'—duplicate before create?"
- "create_task on Project Atlas: title 'Legal review vendor DPA', assignee Maria, due Friday."
Require human "approve" step before bulk creates—especially client-facing commitments.
Anti-patterns
- Auto
create_taskon every noun in transcript—noise floods board - Action items without assignee—orphans by definition
- Creating tasks in personal todo while board stays empty
- Skipping due dates because "we'll remember"
When to use decision_inbox instead
Some meeting outcomes are executive forks, not IC tasks—budget approval, release delay, vendor switch. Route those to add_decision_item / decision_inbox, not create_task.
Rule of thumb: if it needs a C-level choice, inbox; if it needs someone to do work, task.
Bulk extraction without board spam
For long steering calls, AI may propose twelve actions. Batch human triage:
- Now —
create_taskthis sprint - Later — backlog with date
- Not a task — wiki decision only
- Duplicate — link existing task ID from
smart_search
Cap "Now" items per meeting—teams that create fifteen urgent tasks create zero accountability.
Verification loop in standup
Facilitator opens standup with: "Tasks created from Tuesday steering—status by ID?" list_tasks filtered by created date keeps meeting output honest. Items not mentioned twice consecutively get renegotiated or closed—not ignored.
Client-facing action items
External commitments need tighter review: legal phrasing, date feasibility, assignee authority. Never create_task client promises from AI alone—PM confirms against contract and capacity (workload_heatmap on assignee).
Closing the loop by Friday
Set a team norm: every action from Tuesday steering has a task ID by Wednesday standup. Facilitator reads IDs aloud once—culture beats tooling alone.
Slack vs system of record
Slack OK برای ping؛ create_task system of record. rule: اگر در Slack commit شد، bot/reminder «task ID؟» — AI can draft task from thread after human OK.
Action items without owners are wishes. When AI extracts bullets from notes, require name and due date before create_task. Review in standup: orphaned commitments from last week get assigned or explicitly dropped with log_decision so they do not reappear every retrospective.
Extract action items within one hour of meeting end while context is fresh. Tag each draft with meeting date and attendees in the wiki page before create_task so audit trails stay clear six months later.
Run a monthly audit: do AI answers cite task IDs from MCP? If not, tighten prompts to require citations before any stakeholder-facing draft leaves the team.
FAQ — AI meeting action items
Does create_task replace project managers?
No—it records commitments; PMs still prioritize and protect scope.
Can AI assign without asking?
Technically assign_task exists; best practice is confirmed assignee after review.
What about recurring meeting actions?
Use templates or automation rules where configured; still review first occurrence.
How do remote/async teams benefit?
Published wiki + tasks beat timezone-overlap meetings for accountability.
Try it now
Put these patterns on live project data—not slide decks.