You sit through 25 meetings a week, take handwritten notes in 5 of them, send action items from 2, and forget the rest. This tutorial replaces that fog with a 5-minute post-meeting workflow that produces a clean Decisions / Action Items / Parking Lot doc from any Zoom, Meet, or Teams transcript. The audience is managers, PMs, consultants, and founders who run more meetings than they write notes for.
What this covers
A reusable workflow to summarize meetings with AI — transcripts to decisions and action items, with a spot-check loop that keeps you out of trouble when AI misattributes a speaker. We treat AI as the analyst, not the recorder. The recorder is your video platform’s built-in transcription; the analyst is ChatGPT or Claude reading the transcript.
Who this is for
Anyone in 5 or more meetings a week — managers, PMs, consultants, founders, and engineering leads who run standups, planning, and design reviews. If you take perfect notes by hand already, this is overhead. If you regularly walk out of a meeting unsure what was decided, this changes the week.
When to reach for it
When you want to stop writing meeting notes manually but still need accurate decisions and action items. When the input is a long PDF (board deck, memo, regulatory filing) rather than a transcript, the Gemini PDF summarization workflow extracts real value from PDFs more reliably than ChatGPT or Claude. And when the meeting is a research-paper discussion you have to walk into prepared, the 10-minute research-summary workflow gets you conversational on a paper without reading it cold.
When this is NOT the right tool
Highly confidential meetings (compensation, M&A, personnel) where transcripts cannot leave the platform — handle by hand. One-on-one informal chats — overhead exceeds value. Meetings without any clear decisions or actions to extract — the AI will invent some, which is worse than nothing.
Before you start
- Confirm your platform records and transcribes. Zoom, Meet, Teams, Webex all do; some require the host to enable transcription explicitly. Test on one meeting first.
- Set a privacy policy. If transcripts contain customer names, financials, or HR data, use a model with no-training settings (ChatGPT Enterprise, Claude API with no-retention, or self-hosted).
- Pick a single output schema for the team. “Decisions / Action Items (Owner, Due) / Parking Lot / Open Questions” works for 90% of meetings.
- Pre-label speakers if the platform allows. Otter, Fireflies, and Read all support manual speaker labels; that single step removes the #1 AI error.
Step by step
- Record the meeting on your platform of choice (Zoom, Meet, Teams) or via a dedicated app like Otter, Fireflies, Read, or Granola.
- Get the transcript. Most platforms export TXT or VTT; some apps email it within five minutes of meeting end.
- Open ChatGPT or Claude. Paste the transcript with a precise prompt:
Extract from this meeting transcript:
1. Decisions: only items explicitly agreed by 2+ participants
2. Action items: bulleted as "owner — task — due date". If no due date stated, write "TBD".
3. Parking lot: items raised but explicitly deferred
4. Open questions: items raised but not resolved
Do not paraphrase or invent.
If a speaker attribution is unclear, mark "unclear".- Review the output. AI confuses speakers occasionally, and occasionally invents a “decision” that was only a suggestion. Spot-check the 1-2 most important action items by searching the transcript for keywords.
- Send action items to owners individually — a Slack DM or email per owner is far more effective than a group post. The owner who never sees their action does not do it.
- Save the decisions doc somewhere persistent and searchable. Notion, a project channel, or a shared drive — choose one and use it every time. Future-you will search this in 6 months.
First-run exercise
- Pick a low-stakes meeting first — a team standup or a working session — not a board meeting. You will catch the AI’s mistakes faster on familiar material.
- Run the workflow once end-to-end without changing the prompt mid-flight. Save the output.
- Open the transcript and diff against the AI’s output. Count: how many decisions are correct? How many action items have the right owner? How many invented?
- For the second run, change only the prompt — add the rules you wished the first run had. Iterate three times and the prompt is permanent.
Quality check
- Every action item has an explicit owner. “The team will follow up” is not an action item; “Maria will email customer X by Friday” is.
- Every decision is one a participant could repeat. If you would not say it aloud at the next meeting, it is too soft to be a decision.
- Speakers labeled “unclear” are spot-checked before sending. Misattributing an action item to the wrong person creates trust damage that takes weeks to repair.
- Parking lot items have a next-step owner, even if the next step is “Pat brings this to next week’s planning.”
How to reuse this workflow
- Save the schema prompt as a saved instruction or custom GPT. New meeting, same prompt, swap transcript.
- For weekly recurring meetings, also feed in last week’s decisions doc as context. The model picks up on continuity (“this resolves last week’s open question on pricing”).
- Quarterly, audit your decisions docs. Decisions that never got executed, action items with TBD due dates — those are the real meeting failures and they show up in the archive.
Recommended workflow
Record → transcript → AI extract (decisions / actions / parking lot / open questions) → spot-check → distribute actions per owner → archive decisions doc. For on-the-go pre-meeting prep, the ChatGPT voice workflow walks through how to use Voice mode for likely-objection rehearsal before you enter the room.
Common mistakes
- Trusting AI on speaker attribution without check. Misattribution erodes trust faster than missing an item.
- Asking for “meeting notes” — too vague; ask for the four specific outputs.
- Not following up on action items. The summary is the easy part; the chase is where projects move.
- Pasting transcripts of confidential meetings into a consumer chat. Use the right model with retention disabled.
- Letting AI infer decisions that were only suggestions. “Should we…” is not the same as “We will…”.
- Not labeling speakers. The transcription tool’s auto-labels are usually right for 80% of utterances and catastrophically wrong for the other 20%.
FAQ
- Which AI tool is best?: For privacy-sensitive: Claude with no-retention or self-hosted. For convenience: ChatGPT with transcript paste.
- Do I need a dedicated meeting AI?: Not if you have Zoom or Meet transcripts plus ChatGPT. Dedicated tools (Fireflies, Otter, Granola, Read) save time on the export step and add searchable archives across meetings.
- What about meetings in multiple languages?: Modern transcripts handle mixed-language reasonably, but ask the model to summarize in your preferred output language explicitly.
- How long can a transcript be?: Most current models handle a 90-minute transcript in one prompt. Longer than that, summarize in 30-minute chunks then merge.
- Should I share the AI summary or a hand-edited version?: Always edit before sending. Sending raw AI output reads as careless even when it is correct.
- What if no one followed up on the last meeting’s actions?: Start the next meeting with last week’s action list. Public accountability is the only thing that fixes this.