The meeting ended twenty minutes ago. You have a recording, a vague memory of what was decided, and a quiet dread about writing it all up. Sound familiar?
For a long time, that was my Friday afternoon. I’d put off the meeting minutes until the details started to fade — which only made the task harder.
Then I changed one thing. Instead of writing minutes from scratch, I started handing my Microsoft Teams recordings to an AI model and letting it do the first draft. What used to take 40 minutes now takes about 5.
This post is exactly how I do it — no special transcription software, no new subscription. Just the Teams recording you already have and an AI tool you’re probably already paying for.
Quick note: If you want a fully automated version — where an AI joins your call, records, and delivers a summary without any manual steps — I cover that option too. Jump to the automated alternative section if that’s what you’re looking for.
Why AI meeting minutes in Microsoft Teams is harder than it sounds
Here’s something worth noticing: writing meeting minutes isn’t hard because it’s complex. It’s hard because of when it happens.
By the time a meeting ends, your mental energy is mostly spent. Psychologists call this decision fatigue — the more choices and focus a task demands, the less willpower you have left for the next one. Minutes are the “next one.” They land on you precisely when you’re least equipped to handle them.
There’s a second trap, too. The longer you wait, the more the details fade — so the task quietly gets harder the longer you avoid it. Avoidance feels like relief, but it’s actually compounding the cost.
This is why “just be more disciplined” never works. The fix isn’t more willpower. It’s removing the task from the moment you have the least energy — and that’s exactly what handing the first draft to an AI does.
What you need before you start
Good news: you don’t need to install anything new.
Here’s the complete list:
1. A Teams meeting recording
If you record your meetings through Microsoft Teams, the file is automatically saved to your Microsoft 365 account or SharePoint. If you record on a separate device — a phone or tablet — you’ll have an audio file ready to go. Either works.
2. An AI model that accepts audio or long text
You have plenty of options — and the good news is you’re probably already paying for one of these:
- ChatGPT Plus (OpenAI) — the most widely used AI globally. Accepts audio files directly and handles long transcripts well.
- Google Gemini Advanced — strong at summarizing long content. Integrates naturally if you’re already in the Google Workspace ecosystem.
- Claude Pro (Anthropic) — excellent at following detailed formatting instructions. My pick for getting clean, structured minutes.
- Microsoft Copilot — built directly into Microsoft 365. If your organization already uses Teams at scale, this is worth checking first.
- Meta AI — free and increasingly capable. A solid option if you want to test the workflow before committing to a paid plan.
- NotebookLM (Google) — free, and unusually trustworthy for source-based summaries. More on this below.
You don’t need all of them. Pick the one you already use — the workflow in the next section works with any of these.
3. A prompt template
This is the part most AI guides skip — and it’s the reason most people get mediocre output. I’ll give you my exact template below.
Step-by-step: from Teams recording to finished minutes
The process is simpler than most people expect. Here’s exactly how I do it.
Step 1. Get your recording ready
If you recorded through Microsoft Teams, download the file from your Microsoft 365 account or SharePoint. If you used Copilot, even easier — it connects to SharePoint directly, so you can attach the file without downloading anything.
If you recorded on your phone or tablet, open your AI app of choice and attach the file from there. Most major AI apps now support audio attachments on mobile.
One exception worth knowing: if your meeting had many speakers jumping between topics, consider running the audio through a speech-to-text tool first. A clean transcript makes it easier for the AI to separate speakers and follow the thread. For straightforward meetings, though, the raw recording is enough.
Prefer to skip the manual steps entirely?
Everything above assumes you’re starting with a recording you’ve already made. If you’d rather have an AI join your Teams or Zoom call automatically — recording, transcribing, and delivering a summary the moment the meeting ends — Fireflies.ai is the most straightforward option I’ve found for this.
It integrates directly with Microsoft Teams, Zoom, and Google Meet. No file downloading, no manual uploading. The summary lands in your inbox when the call ends.
If you’re running back-to-back meetings and processing recordings manually isn’t realistic, it’s worth a look. They have a free tier to start.
Step 2. Choose your AI model
Any of the models listed above will work. I use Claude for this — mainly because I’ve built a custom meeting minutes format into my Claude Project, so it remembers my preferred structure every time. No re-explaining needed.
That said, most major AI tools now let you save preferences — Claude has Projects, ChatGPT has custom instructions, Gemini has Gems. Once you set your format once, any model can replicate it.
My current workflow: I run the recording through both Claude and NotebookLM, then compare outputs and take the better result. NotebookLM is surprisingly strong at pulling out key points from audio. Having two outputs takes an extra minute but often produces noticeably cleaner minutes.
Step 3. Attach and prompt
Attach your recording or transcript, then send your prompt. This is the step that makes or breaks the output — I’ll give you my exact prompt template in the next section.
Step 4. Review and finalize
When the output comes back, read through it. In my experience, seeing the AI’s summary triggers your own memory of the meeting — details come back that you’d half-forgotten. That’s when I do a quick final edit.
For simple meetings with a clear agenda and two or three speakers, I often use the output as-is. For complex meetings — multiple speakers, topics shifting mid-conversation — a light edit is worth it. Either way, you’re looking at finished minutes in under five minutes rather than forty.
My exact prompt template (copy and adapt)
This is the section most AI guides skip. They tell you to “ask AI to summarize your meeting” — and then wonder why the output looks like a rough draft no one wants to read.
The quality of your AI meeting minutes is directly proportional to the quality of your prompt. Here’s the full template I use, refined over dozens of meetings:
Copy the prompt below. Adjust the fields in brackets to match your meeting.
“Using the uploaded audio or text file, create meeting minutes with the following requirements:
Format and Output
1. Deliver the output as a Word document (.docx).
2. Keep the writing style concise and professional.
3. Correct any typos or misheard terms based on the context of the meeting topic.
4. Add page numbers at the bottom of each page.
Document Structure
5. Write an Executive Summary at the top. Number each key point for easy scanning.
6. Organize the meeting content by agenda item, with sufficient detail under each.
7. Include a Follow-up / Action Items section at the end.
Header Information
8. Meeting title: [insert title]
9. Date and time: format as — May 13, 2026 (Wed) 13:00
10. Location: [insert briefly]
11. Author: [insert name and title]
12. Attendees: group by organization or department if identifiable from the recording.
Design
13. Use a clean, readable font throughout.
14. Make the meeting title text slightly lighter in color than the Meeting Minutes heading above it.
15. Omit the signature/approval block at the bottom of the document.
After the file: Separately (not in the Word document), provide a 3-line summary I can paste into a chat or email when sharing the file.”
A few notes on how I use this:
I don’t use the full prompt every time. For a quick internal check-in, I strip it down to the essentials — structure, summary, action items. The full version is for formal meetings where the output goes to senior stakeholders.
The line about correcting typos and misheard terms (point 3) is one of the most underrated instructions you can give. AI will mishear technical terms, product names, or people’s names. Telling it to self-correct based on context saves you a manual proofread.
The 3-line summary at the end is something I added after realizing I was rewriting a summary every time I shared the file. Now it comes automatically.
Adapt it to your language. I work across two languages — English and Korean — and I use versions of this prompt in both. The AI handles either without issue. Whatever language your meetings run in, write your prompt in that language. The output will follow.
ChatGPT vs Gemini vs Claude vs NotebookLM: which AI is best for meeting minutes?
I’ve run this workflow through all of them. Here’s what I’ve found — not as a benchmark test, but as someone who uses these tools in real meetings, across two languages, week after week.
ChatGPT (OpenAI)
ChatGPT is reliable and consistent. It follows instructions well and produces clean output. The limitation I’ve noticed for AI meeting minutes specifically: it tends toward brevity. Give it a long, detailed meeting and it will summarize — sometimes a little too aggressively. If your minutes need to capture nuance or complex decisions, you may find yourself prompting it to expand.
Best for: Quick internal meetings where a tight summary is exactly what you want.
Google Gemini
Gemini’s strength is flexibility. Of all the models I’ve used, it handles format switching most smoothly — I’ve taken a meeting summary in Gemini and converted it into a presentation outline without starting over. If your workflow goes beyond minutes into decks or reports, Gemini transitions well.
The tradeoff: the output can feel slightly rigid in tone. It does the job, but the writing doesn’t always flow naturally. Worth noting if your minutes go to senior stakeholders who care about how things read.
Best for: Workflows where you need the same content in multiple formats.
Claude (Anthropic)
Claude is my primary tool for this task — mainly because of how well it handles structured, detailed prompts. Give it the full template from the previous section and it follows every instruction without losing any of them halfway through. The output tends to read more naturally than the others.
The other reason I use Claude: Projects. I’ve saved my preferred meeting minutes format directly into a Claude Project, so I don’t re-explain my preferences every session. It remembers.
Best for: Formal meetings where output quality and formatting precision matter.
NotebookLM (Google)
NotebookLM works differently from the others — and that difference is actually its biggest strength for this use case. It only references what you’ve uploaded. It won’t hallucinate details, fill gaps with assumptions, or drift into general knowledge. What’s in the recording is what goes into the minutes — nothing more.
This makes it unusually trustworthy for meeting documentation. I often run the same recording through both Claude and NotebookLM and take the better output. They each catch things the other misses.
Best for: Meetings where accuracy is critical and you can’t afford fabricated details.
Quick comparison:
| Tool | Strength | Watch out for |
|---|---|---|
| ChatGPT | Clean, consistent output | Can over-summarize |
| Gemini | Format flexibility | Tone can feel stiff |
| Claude | Structured, natural output | Best with detailed prompts |
| NotebookLM | Stays strictly on source | Less flexible formatting |
Want to skip the tool selection entirely?
If choosing between models sounds like another decision you don’t need, Fireflies.ai removes that layer — it handles recording, transcription, and summary automatically, and delivers a structured output without you touching any of the tools above. Worth considering if automation matters more than customization.
Common mistakes — and how to avoid them
The workflow is simple, but a few habits will make the difference between output you use as-is and output you spend ten minutes fixing.
Mistake 1. Uploading one long recording for an entire day
This is the most common one. A two-hour recording covering three different topics will produce muddled minutes — the AI loses the thread as the conversation shifts.
The fix is simple: record in segments.
Two rules that work well in practice:
- Break recordings at the one-hour mark. Process each segment separately, then combine the outputs.
- Split by topic when you can. If your meeting moves from budget review to project update to HR matters, those are three separate recordings. The minutes for each will be noticeably more accurate.
This takes about five extra seconds of discipline during the meeting. The payoff in output quality is significant.
Mistake 2. Skipping the prompt and just uploading the file
“Summarize this meeting” will get you a summary. It won’t get you structured minutes with an executive summary, action items, and proper formatting. The prompt is where the output quality lives. Use the template above, even in a stripped-down version.
Mistake 3. Treating the first output as final
AI-generated minutes are a first draft, not a finished document. Read through the output once — you’ll find that reviewing it triggers your own memory of the meeting. That’s when you catch anything the AI missed or slightly misrepresented.
Mistake 4. Ignoring speaker identification
If your meeting has more than three or four speakers and speaker clarity matters for the record, run your audio through a speech-to-text tool first. A clean, speaker-labeled transcript gives the AI much more to work with.
Mistake 5. Using the wrong language in your prompt
Write your prompt in the same language your meeting was conducted in. If the meeting was in Korean, prompt in Korean. If it was in English, prompt in English. Keep it consistent.
Is it worth it? The honest ROI
Let me be direct.
If you run or attend meetings regularly — and especially if you’re responsible for documenting them — this workflow will save you meaningful time every week. Not “eventually” or “in theory.” From the first meeting you try it on.
Here’s a rough comparison based on my own experience:
| Without AI | With AI | |
|---|---|---|
| Writing from memory | 30–40 min | — |
| Reviewing AI draft | — | 3–5 min |
| Formatting and sending | 10 min | 2 min |
| Total | 40–50 min | 5–7 min |
That’s roughly 40 minutes returned to you per meeting. If you document two meetings a week, that’s over an hour and a half back — every week.
The more important shift, though, isn’t the time. It’s the mental load.
Writing minutes from scratch requires you to reconstruct a conversation you just lived through, while your energy is already spent. Removing that task from your plate — and handing the first draft to an AI — means you show up to the next thing fresher. That compounds over a week, a month, a quarter.
Get started today
What you need:
- A meeting recording (Teams, phone, tablet — any format works)
- One AI tool you already use
- The prompt template above
That’s the whole system. No new software. No extra cost beyond what you’re probably already paying.
If you found this useful, the next post takes a broader look at the AI landscape — not just language models, but the full toolkit: which AI generates the best images, which handles video, which is built for writing, and which is worth trusting for research and analysis. A practical map of what each tool actually does best →
Tools mentioned in this post
- ChatGPT Plus — most widely used, reliable across tasks
- Google Gemini Advanced — best format flexibility
- Claude Pro — my pick for structured, detailed work
- NotebookLM — free, and uniquely trustworthy for source-based summaries
- Fireflies.ai — if you want a fully automated meeting assistant that joins, records, and summarizes for you (affiliate link — I earn a commission if you sign up, at no extra cost to you)
Not sure which AI to use for different tasks? Read: The Right AI for the Right Job