Capture the meeting with a note-taker (Zoom, Teams, Google Meet or Otter), have AI extract each action as owner, task, due date and project in a fixed format, send that list to the meeting owner for a two-minute approval, then push the approved items into your task tool through a native integration or a Zapier or Make workflow.
Transcription is no longer the weak point; the tools are good at it. What breaks this workflow is how people talk. "We should look into that" has no owner, "by Friday" means different things on Monday and Thursday, and half of what sounds like an action is an idea nobody committed to. Fix those with a few meeting habits and strict prompt rules, and keep a human approval in the middle, and the tasks that land are ones people recognise and do.
What your note-taker already does before you add anything
Check this first, because for some teams the built-in feature is enough.
| Tool | What it produces | Where actions go | What you need |
|---|---|---|---|
| Microsoft Teams, Facilitator | Real-time AI notes in a Loop page that attendees can edit | Captured tasks can be synced to Planner through its follow-up tasks ("Accept to sync"); some task commands are in preview | A Microsoft 365 Copilot licence |
| Google Meet, Take notes for me | A Doc with Summary, Decisions, Next steps and Details sections, saved in the organiser's Drive | Next steps listed in the Doc; you move them yourself | An eligible Workspace edition or Google AI plan |
| Zoom meeting summary | Summary with next steps from the transcript (the features once branded AI Companion) | Summary email and Zoom's own notes; workflows vary by plan | An eligible paid Zoom Workplace plan |
| Otter.ai | Transcript, summary and action items | Integrations with some task tools, plus Zapier | Pro from $8.33 a user a month billed annually |
If your team lives in Teams and Planner and already pays for Copilot, try Facilitator's own task sync before building anything. Microsoft's Facilitator page describes the current behaviour. Everyone else, and anyone whose tasks live in Asana, Trello, ClickUp or Monday, needs the extraction and push steps below. The comparison of AI meeting note-takers for small teams helps if you haven't picked one yet.
Run the meeting so the AI can find the actions
Ten seconds of discipline per action makes extraction reliable. Three habits:
- Say actions as owner, verb, date. "Maya will send the revised press release to the client by Thursday the 8th" extracts perfectly. "Let's get the release sorted" doesn't.
- Close with a 90-second recap. The chair reads the actions back. Now every action appears twice in the transcript, once in discussion and once cleanly, and the AI can prefer the recap version.
- Name who's in the room. Transcripts label speakers badly when people join from one meeting room. A quick round of names at the start helps the tool, and helps the extraction match first names to your team list.
Here's a realistic snippet of what poor and good phrasing look like in a transcript, from an illustrative PR agency client call:
POOR [14:02] Jonah: We probably need to get the spokesperson briefed before
the trade press thing.
GOOD [14:03] Jonah: So, action for Priya: brief Dr Okafor for the trade press
interview, by Monday the 12th.
GOOD [27:40] Chair recap: Priya, spokesperson brief, Monday 12th. Client
to confirm embargo time by Thursday. Jonah, media list update,
end of this week.An extraction prompt that returns clean tasks
Whether you paste a transcript into an assistant by hand or run it through an automation, use the same prompt. The rules matter more than the wording.
You are extracting action items from a meeting transcript.
Meeting date: {{meeting_date}} (use this to convert relative dates).
Our team: {{team_list with first names and emails}}
Client/project: {{project_name}}
Return JSON: a list of actions, each with:
- task: starts with a verb, under 12 words
- owner: a first name from Our team, or CLIENT, or UNASSIGNED
- due_date: YYYY-MM-DD, or null if no date was agreed
- source_quote: the exact words that created the action, with timestamp
- confidence: high if an owner and task were stated explicitly, else low
Rules:
- Only include commitments. Ideas, options and "we could" are not actions.
- If the recap and the discussion disagree, use the recap.
- "End of week" means the Friday of the meeting's week. "Next week" with no
day means null, not a guess.
- Merge duplicates (same task, same owner) into one action.
- Never invent an owner. If unclear, UNASSIGNED.An illustrative output for the call above (meeting date Tuesday 6 October):
[
{"task": "Brief Dr Okafor for trade press interview", "owner": "Priya",
"due_date": "2026-10-12", "confidence": "high",
"source_quote": "[14:03] action for Priya: brief Dr Okafor... by Monday the 12th"},
{"task": "Confirm embargo time", "owner": "CLIENT",
"due_date": "2026-10-08", "confidence": "high",
"source_quote": "[27:40] Client to confirm embargo time by Thursday"},
{"task": "Update media list", "owner": "Jonah",
"due_date": "2026-10-09", "confidence": "high",
"source_quote": "[27:40] Jonah, media list update, end of this week"},
{"task": "Explore podcast placements for Q1", "owner": "UNASSIGNED",
"due_date": null, "confidence": "low",
"source_quote": "[19:15] we could look at a few podcasts for Q1"}
]What you'd fix at approval: the fourth item is exactly the kind of idea the rules say to exclude, and the model included it with low confidence anyway. That's fine; the approval step exists for it. Delete it, or give it an owner if someone really should explore it. Also check the dates: Thursday the 8th and Friday the 9th are right for a meeting on Tuesday the 6th, and that arithmetic is worth a glance every time because it's where models slip.
The two-minute approval step
This is the part people want to skip, and the reason the workflow survives. Without it, tasks nobody agreed to land in people's lists, a few get ignored, and within a month the team stops trusting anything the automation creates.
The simplest reliable version uses a spreadsheet as the holding area:
- The extraction step writes each action as a row in a "Pending actions" sheet, with the meeting name, a link to the notes, and an empty Approved checkbox.
- The meeting owner gets one message: "6 actions from the Tuesday client call are waiting, 1 low-confidence." They open the sheet, fix owners and dates, delete non-actions and tick Approved.
- A second workflow watches for ticked rows and creates the tasks.
A message-based version works too: a Slack or Teams message listing the actions, with the owner replying "approve" or editing. The sheet is easier to build and easier to audit.
Pushing approved actions into Asana, Trello or Planner
Map the fields once. Here's the mapping an illustrative agency used for Asana; Trello and Planner have direct equivalents.
| Extracted field | Task tool field | Notes |
|---|---|---|
| task | Task name | Prefix with the client code, e.g. "[RTL] Brief Dr Okafor" |
| owner | Assignee | Look up the email from the team list; skip CLIENT rows |
| due_date | Due date | Leave blank if null rather than inventing one |
| project | Project | One project per client keeps workload views accurate |
| source_quote + notes link | Description | Lets the assignee see exactly what was agreed |
| (fixed) | Tag "from-meeting" | So you can measure how these tasks perform |
Build it in Zapier or Make: trigger on a new approved row, look up the assignee, create the task, then mark the row "Created" with the task link so it can never be created twice. The tutorial on adding AI steps to Zapier covers the extraction side if you want the whole thing automated from the transcript onwards.
A quick cost sum on Zapier: triggers and filters don't count as tasks; each action step does, and an AI by Zapier step uses 1, 3 or 5 tasks depending on the model tier. Say each meeting runs the AI step at 3 tasks, writes 6 rows (6 tasks), sends the approval message (1) and later creates 5 approved tasks (5). That's 15 tasks a meeting. At 50 meetings a month, 750 tasks, which is exactly the allowance on Zapier Professional ($19.99 a month billed annually). Busier teams need a higher tier, or can batch rows into one step. Make, which prices by credits from about $9 a month, is often cheaper at this kind of volume.
Recurring actions, multi-step work and actions with conditions
Three kinds of action trip up a simple one-row-one-task pipeline. Decide in advance how each is handled and add the rule to the prompt.
- Recurring actions. "Jonah sends the coverage report every Friday" isn't one task; it's a standing commitment. Have the prompt tag these as RECURRING and route them to the approver, who sets up a repeating task by hand once. Creating a fresh task every time the call repeats it gives you duplicates within a fortnight.
- Multi-step work. "Draft the launch release, get legal sign-off, send to the wire service" is three steps with two owners. Ask for one parent task with the steps listed in the description, owned by whoever drives it. Splitting into three tasks with dates the meeting never agreed invents a plan.
- Conditional actions. "If the client approves the budget, Maya books the photographer" shouldn't become a task for Maya yet. Tag it WAITING ON, with the condition in the description, and create it only when the condition is met.
Those three tags cover most of what goes wrong in agency meetings, and they keep task lists honest: a list full of things people can't start yet teaches them to ignore it.
Meetings with no note-taker: typed notes and voice memos
Plenty of useful meetings happen in a room, on a phone call or on a site visit where nothing records. The pipeline still works as long as you feed it text.
Consider a marketing agency's quarterly planning session with a client, held at the client's office. The account director opens a voice memo on their phone in the taxi afterwards and talks for 90 seconds: "Actions from the planning session. Me: send the Q1 channel plan by the 20th. [Colleague]: brief the designer on the new landing page this week. Client: confirm the budget split between paid social and search by next Wednesday. Parked: the podcast idea, revisit in January." The phone's transcription, or Otter, turns it into text, and the same extraction prompt produces four clean rows, with the podcast item tagged as not an action because it was explicitly parked.
This is often more accurate than a full transcript, because the person dictating has already done the judgement: only real commitments get said out loud. Typed notes work the same way if whoever takes them writes each action on its own line as owner, verb, date.
When the wrong person gets the task
The most common real-world failure is mis-assignment, and it rarely comes from the prompt. A realistic example: in a hybrid meeting, three people join from one meeting-room laptop. The transcript labels all three as "Meeting Room 2", and the chair's recap says "you'll send the proofs, and you'll call the printer". The extraction, correctly following its rules, returns UNASSIGNED for both. On a busy day the approver, reading quickly, assigns both to the person whose name appears most in the transcript. One task sits untouched for a week before anyone notices.
Fixes that work: in hybrid meetings, ask people in the room to say their name before taking an action; make the approver's message list UNASSIGNED items at the top in bold rather than mixed in; and add a rule that a "you" with no name is always UNASSIGNED, never guessed. If mis-assignments keep appearing in your monthly check, look at the meeting set-up before you touch the prompt.
A nine-person PR consultancy's weekly client calls
Here's how the numbers can work at a nine-person PR consultancy, with illustrative figures. It runs 14 client status calls a week on Google Meet plus two internal planning meetings.
Before: after each call an account executive typed up notes and actions, emailed them round and added tasks to Asana. About 20 minutes per call, so roughly 4.5 hours a week. Actions from Friday calls often waited until Monday, and client-owned actions were rarely chased.
Set-up (one day, spread over a week): Take notes for me switched on for client calls, with clients told at the start of each call and in the meeting invite. A workflow picks up each new notes Doc from the Meet folder in Drive, sends the text through the extraction prompt, and writes rows to the Pending actions sheet. Account leads approve; a second workflow creates Asana tasks.
After four weeks (illustrative):
- Approval takes about 3 minutes a call, so roughly 45 minutes a week instead of 4.5 hours.
- About 6 actions extracted per call; account leads edit roughly 1 in 5 (usually the owner or date) and delete about 1 in 10 (ideas, not commitments).
- Client-owned actions go into a follow-up email the same afternoon instead of the next morning.
- Early problem: the two internal planning meetings produced 25-30 "actions" each because they're brainstorms. The fix was to exclude them from the workflow; brainstorms need a person to decide what's actually happening.
Turning client actions into a same-day follow-up email
For client-facing meetings, the most visible benefit is the recap. Once actions are approved, the AI drafts the follow-up from the approved rows only, never from the raw transcript, so nothing unapproved slips into an email to the client.
Draft a short follow-up email to {{client_contact}} after today's status call.
Use only the approved actions below. Group into "We'll do" and "Could you".
Warm but brief, no recap of the discussion, no new commitments.
Approved actions: {{approved_rows}}An illustrative draft:
Hi [client first name],
Thanks for the time this morning. Quick note of where we landed:
We'll do
- Brief Dr Okafor for the trade press interview (by Mon 12 Oct)
- Update the media list for the product launch (by Fri 9 Oct)
Could you
- Confirm the embargo time by Thu 8 Oct
Speak next Tuesday,
[your name]The account lead reads it, adjusts the tone if needed, and sends. Their check is quick because every line maps to a row they already approved. A marketing agency would do the same after campaign reviews, with "Could you" typically holding approvals of creative and budget sign-offs, which are the client actions that stall campaigns most often.
Signs the workflow is drifting
Automations like this decay quietly. Put four numbers on a monthly check, all easy to pull from the Pending actions sheet and the "from-meeting" tag:
- Edit rate at approval. If more than about a third of actions need edits, the prompt or the meeting habits need attention. Look at which field is edited most.
- Deleted as non-actions. Rising deletions usually means a new meeting type (a brainstorm, a workshop) has crept into the workflow.
- UNASSIGNED count. A steady trickle is normal. A spike means someone has stopped naming owners in meetings.
- Completion by due date for "from-meeting" tasks compared with other tasks. If they're done less often, people don't feel they own them, which points back to approval being rushed.
Also check the plumbing: a notes Doc that never arrived, or a workflow that failed on a long transcript, won't announce itself. A daily count of meetings held against rows written catches it.
Consent, privacy and meetings you shouldn't record
Tell everyone when a meeting is being transcribed, in the invite and again at the start, and give external participants a real chance to object. Some meetings shouldn't go through a note-taker at all: HR conversations, sensitive client matters, anything under legal privilege. For those, a person writes the actions by hand, and they can still go through the approval sheet if you want them tracked the same way. Before you switch transcription on for client calls, read whether AI note-takers are safe for client calls, and if you're on Teams, the settings covered in Copilot in Teams meetings: recaps, actions and privacy.
If your team's tasks are split across tools, fix that first: the tutorial comparing ClickUp, Asana and Monday AI features can help you settle on one, and the meeting workflow gets much simpler once every action has a single place to land.
Follow-up questions about meeting-to-task automation
Do I need to record the meeting for this to work?
You need text, not necessarily a recording. A transcript from a note-taker is easiest, but typed notes work if whoever takes them writes actions as owner, verb and date. For meetings where recording isn't appropriate, dictate a one-minute voice memo straight after and run that through the same extraction prompt.
What about actions for people outside the business, like clients?
Keep them, but don't create tasks in your own tool for them. Tag them as client actions and put them in the follow-up email instead, with a task for your account lead to check they happened. That way your task list only holds work your team owns.
Should the AI assign tasks to people directly?
Only after a person approves the list. Assignment is where extraction goes wrong most often, because conversation is full of 'we should' and 'someone needs to'. The two-minute approval step lets the meeting owner fix owners and dates before anyone's task list fills with things they never agreed to.
Further reads
- Zoom AI Companion for Small Teams: Summaries and Privacy Settings — Set up Zoom's summaries and privacy controls properly.
- How to Update Your CRM Automatically After Every Sales Call — The same pattern for sales calls, into your CRM.
- How to Stop Zapier and Make Automations Breaking Silently — Get alerted when the task workflow fails quietly.
- AI vs Human Transcription: Which Is Worth Paying For? — When a human transcript is worth paying for.
- Which Agency Tasks AI Handles Well, and Which It Doesn't — Other agency jobs worth handing to AI.
- How to Get a Weekly Business Summary Emailed to You by AI — Roll completed actions into a weekly summary.
- Best AI Tools for Small Event Planning Businesses — Eight AI tools for small event planning businesses, ranked by the hours they give back, with prices, first uses and the two to handle with care.
- A Wedding Planner's AI Workflow From Enquiry to Final Timeline — One illustrative planner's year with AI, stage by stage: consultation notes, proposal, suppliers, guests, confirmations and the wedding-day timeline.
- Which Event Planning Tasks Should You Hand to AI First? — Twelve event planning tasks scored for AI, the five to hand over first with worked examples, and the ones to keep even though AI will attempt them.
- AI Session Notes and Follow-Ups for Coaches: A Weekly Workflow — A week-shaped routine for coaches: capture sessions with consent, turn transcripts into your own notes, send same-day follow-ups and review clients on Friday.
- How Advisers Use AI to Prepare for Annual Client Reviews — A four-week countdown for review prep with AI: what it drafts, what the adviser checks, and the prompts that keep figures and advice in human hands.
- How Architects Use AI for Fee Proposals, Briefs and Paperwork — How a small practice uses AI for briefs, fee proposals, variation letters and site reports, with prompts, sample outputs and the parts AI must leave alone.
- How Small Accounting Firms Use AI Day to Day: Real Examples — A five-person practice's working week with AI, day by day: the tools, the prompts, what came back, and what the team corrected.
- Can Lawyers Use AI Note Takers in Client Meetings? — When an AI note taker is fine in a client meeting, when it isn't, the consent wording, and how to turn a transcript into an attendance note.
- How Much Time Can AI Save a Solo Consultant Each Week? — A task-by-task estimate of the hours AI saves a one-person consultancy, three example weeks, the hidden costs and a two-week way to measure yours.
- Can AI Help a Small Architecture Practice? Where It Saves Time — Follow an illustrative client review pack, then test AI on briefs, meeting actions, drawing comments and early design options.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Microsoft Support, Facilitator in Microsoft Teams meetings; Google Meet Help, Take notes for me; Zoom support pages on meeting summaries; Otter.ai pricing and integrations pages; Zapier pricing and task-counting documentation. Checked September 2026.