How to use AI to turn a recorded team meeting into a clean set of action items and decisions without a minute-taker

A 1-hour Zoom call produces 6,000–9,000 words of transcript — and most of it sits unread in your account until someone needs to remember what was decided. This post walks you through a three-step workflow to turn that transcript into a clean action-item list using AI tools you likely already have open. The whole process takes under 10 minutes per meeting, and for recurring calls it drops to under 5 once you've saved a prompt template.

What you need before you start

Zoom{:target="_blank"}, Google Meet{:target="_blank"}, Fathom{:target="_blank"}, or any recording tool that exports a speaker-labeled transcript — you need the transcript file, not just the video. Zoom's native AI Companion is included free with paid Zoom accounts and can auto-generate summaries, but you'll still want the raw transcript for the workflow below. Fathom's free tier (as of 2025) covers recording, transcription, and highlights without a paid plan.

ChatGPT{:target="_blank"} (GPT-4o) or Claude{:target="_blank"} (3.5 or 3.7 Sonnet) — either works. GPT-4o handles up to 128,000 tokens; Claude 3.5 and 3.7 Sonnet support up to 200,000 tokens. Both easily handle a full 1-hour transcript in a single paste. ChatGPT pricing{:target="_blank"} starts at $20/month for Plus; Claude pricing{:target="_blank"} starts at $20/month for Pro. The free tiers of both can process transcripts but may have rate limits during heavy use.

Time required: 5 minutes for a one-off meeting once your prompt is built. Under 10 minutes including transcript export. First-time setup — building and saving your prompt template — takes about 20 minutes.

Skill level: No technical background needed. You need to know how to copy and paste text and how to export a file from your recording tool.

---

Step 1: Get a clean, speaker-labeled transcript out of your recording tool

1. Open your recording in Zoom, Fathom, Google Meet, or whichever tool captured the call.
2. Locate the transcript export option. In Zoom, go to the Zoom web portal{:target="_blank"}, find the recording, and select "Download" → "Transcript (VTT or plain text)." Choose plain text — it's cleaner to paste.
3. Check that speaker names are labeled in the export. You'll see something like Marcus: Let's move the launch to the 15th. If names are missing or appear as "Speaker 1" and "Speaker 2," stop here and fix it first (see the failure scenarios below).
4. Copy the full transcript text to your clipboard.

The speaker-labeling step is not optional. The single most common failure mode when pasting transcripts into AI is the model assigning tasks to the wrong person — or hallucinating an assignment entirely — because it couldn't tell who said what. Spend 60 seconds verifying names before you paste.

---

Step 2: How to summarize a meeting transcript with AI and get structured output

This is where most small business owners get it wrong — they paste the transcript and type "summarize this." The output is a paragraph of prose that buries decisions inside context. You need structured output with explicit formatting instructions.

1. Open a new chat in ChatGPT or Claude.
2. Paste the following prompt at the top of your message, then paste the full transcript immediately below it — no blank line between the prompt and the transcript.

Prompt template — copy and use directly:

You are processing a business meeting transcript for a small business owner. Your job is to extract three things with precision:

1. Action items — List every task that was assigned or agreed to. Format each one as:

• Who: [Name of the person responsible]
• What: [Specific task, described in plain language]
• By when: [Deadline if stated; write "Not specified" if no date was given]
• Priority: [High / Medium / Low — infer from urgency language in the transcript]

2. Decisions made — List every decision that was reached and agreed to by the group. One sentence per decision. Include who made or confirmed each decision if clear from the transcript.

3. Unresolved items — List every question, discussion point, or potential decision that was raised but NOT resolved by the end of the meeting. These need follow-up.

Rules:

• Do not invent tasks. Only include items explicitly mentioned in the transcript.
• If a task was raised but no one accepted ownership, flag it under Unresolved Items, not Action Items.
• Keep descriptions brief and specific. No summaries of conversation, no filler.

Transcript begins below: [PASTE TRANSCRIPT HERE]

3. Send the message and wait for the output. A 9,000-word transcript processes in under 30 seconds on either platform.
4. Review the three sections — Action Items, Decisions Made, and Unresolved Items — against your own memory of the call. The AI will occasionally miss a soft commitment ("I'll look into that") or flag something as decided when it was only proposed. Spot-checking takes 2–3 minutes.

The three-part output structure — action items separated from decisions, both separated from unresolved items — is the detail that makes this actually useful. Teams consistently follow up on tasks more reliably when they have a written list reviewed within 24 hours of the meeting. Blurring decisions and open questions into the same list is what kills that follow-through.

---

Step 3: Turn the AI output into a task list your team will actually use

The AI output is structured but not yet in your task manager. Here's how to move it there without re-typing everything.

1. Copy the Action Items section from the AI output.
2. Open a second prompt (new message in the same chat) and type:

Reformat the action items above into a simple table with four columns: Assignee | Task | Due Date | Priority. Use "TBD" for any due date that was not specified. Keep task descriptions under 15 words.

3. Copy the table output into your task manager — Asana{:target="_blank"}, Trello{:target="_blank"}, Notion{:target="_blank"}, a shared Google Sheet, or whatever your team uses. Most can accept pasted text or a markdown table.
4. Send the Unresolved Items list directly to whoever owns each topic — usually by email or Slack — with a note that these need a decision before the next meeting.

The two-step prompt sequence (extract first, reformat second) produces more reliable output than asking for everything in a single prompt. One job at a time gives the model cleaner context to work with.

---

What to do when you don't have a transcript yet

Some recordings — older Zoom calls, phone call recordings, voice memos — don't have an auto-generated transcript. You have two options:

Option A: Use Fireflies.ai. Fireflies.ai{:target="_blank"} has a free tier that transcribes uploaded audio files and generates an AI summary. Its Pro plan is $18/user/month as of early 2026{:target="_blank"} and adds CRM and task manager integrations (Asana, Trello, HubSpot). Upload your audio file, let it process (usually 5–10 minutes for a 1-hour file), export the plain-text transcript, and run it through the prompt workflow above.

Option B: Use Whisper via ChatGPT. OpenAI's Whisper{:target="_blank"} transcription model is built into ChatGPT's voice and file features on the Plus plan ($20/month). Upload the audio file directly to a ChatGPT conversation and ask it to transcribe. Quality is high, though speaker labeling is not always reliable on Whisper output — review before pasting into your action-item prompt.

---

How to get action items from a Zoom recording: dedicated tool vs. general AI

Fathom{:target="_blank"} and Fireflies.ai{:target="_blank"} both connect directly to your calendar and join meetings automatically. Otter.ai{:target="_blank"}'s Business plan runs $20/user/month as of early 2026{:target="_blank"} and includes CRM integrations, but the free plan caps at 300 minutes of transcription per month and 3 imports — tight for most active small businesses.

Here's the trade-off: dedicated tools reduce friction by automating the transcript capture step, but their built-in summaries often miss context-specific decisions that are unique to your business — a project code, a client nickname, a pricing exception your team understands but the AI doesn't. The general AI workflow (ChatGPT or Claude plus a good prompt) produces more accurate action items because you're adding the context the model needs.

The honest answer is: use a dedicated notetaker to capture and transcribe automatically, then run the transcript through the prompt workflow above rather than relying on the tool's native summary. You get the automation of the former and the accuracy of the latter.

---

Save the workflow: building a reusable meeting processing routine

1. Save the prompt template as a text file or note titled "Meeting Processing Prompt." Most teams put it in Notion or a shared Google Doc.
2. Create a standing task in your project manager: every Monday (or whatever your recurring meeting day is), whoever runs the meeting exports the transcript and pastes it into ChatGPT using the saved prompt.
3. Set a rule: action items get entered into the task manager within 2 hours of the meeting ending. Unresolved items get a reply-all email or a Slack message the same day.

For recurring meetings — weekly standups, client check-ins, operations reviews — add one line to the prompt template:

Context: This is a weekly [team standup / client check-in]. Previous action items that appear in the transcript as "updates" have already been completed unless otherwise noted. Focus only on new commitments and new open questions.

That single addition eliminates the most common confusion on recurring meeting transcripts, where completed tasks from last week get re-listed as new assignments this week.

---

When something goes wrong

Symptom: The action items list includes tasks assigned to the wrong person, or mentions "the team" without naming anyone. Root cause: The transcript has missing or inconsistent speaker labels — "Speaker 1" instead of a name, or a name that changes mid-transcript. Fix: Before pasting into AI, do a find-and-replace in the transcript to standardize names. Replace "Speaker 1" with the actual person's name. Even a quick manual pass on the first five lines often resolves 80% of attribution errors.

Symptom: The AI lists a decision under Action Items or buries it in the summary paragraph instead of the Decisions section. Root cause: The decision was phrased tentatively in the meeting ("I think we're going with option B, right?") so the AI flagged it as uncertain. Fix: Add this line to your prompt: When in doubt about whether a decision is final, flag it under Unresolved Items rather than Decisions Made. This biases toward caution, which is the right call — you'd rather review a false alarm than miss an actual open question.

Symptom: The output is a wall of prose instead of the structured three-part format you asked for. Root cause: The transcript was pasted before the prompt, so the model read the content before it read the instructions. Fix: Always paste the prompt first, transcript second, with the phrase "Transcript begins below:" as the divider. Instruction-first order matters for structured output.

---

What to do next

Build your prompt template this week and run it against your last recorded meeting — even a completed one — to test the output quality before you rely on it for a live call. The prompt above requires no customization for most meetings, but adding one line of business context ("This is a weekly call with our production team; we manufacture custom furniture") improves specificity noticeably.

If you want to extend this workflow into your CRM or project manager automatically, read about connecting AI outputs to your task tools without code.

---

FAQ

How do I get action items from a Zoom recording if I don't use AI Companion? Export the transcript from the Zoom web portal{:target="_blank"} as plain text (available on paid Zoom plans), then paste it into ChatGPT or Claude using the prompt template in Step 2 above. Zoom's AI Companion auto-summary is included free with paid accounts but produces less structured output than running the transcript through a specific prompt — treat it as a rough draft, not a finished list.

Is there a way to get meeting action items without paying for anything? Yes. Fathom's free tier (as of 2025) transcribes and records meetings at no cost. Claude and ChatGPT both have free tiers that can process transcripts, though rate limits apply during peak hours. The full workflow described above — Fathom for capture, free ChatGPT or Claude for processing — costs $0 per month for most small business meeting volumes.

What does Otter.ai cost, and is it worth it for a small team? Otter.ai's free plan allows 300 minutes of transcription per month and 3 file imports — enough for roughly 3–4 short meetings. The Business plan runs $20/user/month as of early 2026, which adds integrations with Salesforce, HubSpot, and productivity tools. The cost is hard to justify for teams that primarily need transcripts and action items; Fathom plus a general AI tool covers that workflow for significantly less. Otter's Business plan earns its cost if your team needs CRM auto-sync and you're already paying for those integrations.

Will the AI make up action items that weren't in the meeting? The risk is real but manageable. Models like GPT-4o and Claude 3.5/3.7 Sonnet hallucinate less on structured extraction tasks than on generation tasks, but they can infer a commitment from a vague statement ("I'll think about that") that the speaker didn't intend as a firm task. The prompt template above includes the instruction "Do not invent tasks. Only include items explicitly mentioned in the transcript." Combined with a 2–3 minute spot-check against your own memory of the call, hallucination risk drops to negligible.

How much time does this workflow actually save? A 1-hour meeting typically takes 30–45 minutes to manually summarize into a usable action list — longer if the note-taker wasn't on the call. The AI workflow, including transcript export and prompt processing, runs in under 10 minutes the first time and under 5 minutes for recurring meetings with a saved template. For a team running three 1-hour meetings per week, that's roughly 2–3 hours per week recovered. At a $50/hour opportunity cost, that's $5,000–$7,800 per year in reclaimed time — from a workflow that costs nothing beyond tools you likely already pay for.