The meeting ends at 11. By 11:40 you are staring at a blank document, trying to remember who agreed to what. Was it Sarah who said she would send the pricing sheet, or was it Tom? Did the client approve the new timeline, or did they just nod while looking at their phone? You write down what you think happened, send it round, and three days later someone replies: "That is not what I agreed to."
That gap, between what a meeting decided and what anyone remembers, is the whole reason AI meeting notes exist. Used properly, they save you the twenty minutes after every meeting. Used badly, they produce a confident-looking document full of decisions nobody made. This guide covers the honest version: how to use AI to write meeting notes and action items, the exact prompts to copy, a worked example, and the two-minute check that stops the mistakes.
Can AI write meeting notes and action items for you?
Yes. If you give an AI tool a recording or a transcript, it can turn an hour of rambling into a one-page summary with decisions and action items in under a minute. If you paste in rough notes instead, it can clean them into the same format. What it cannot do is decide what was actually agreed when the meeting itself was unclear, and it will sometimes fill that gap with a guess that looks like a fact. Treat the output as a draft, not the record, and the tool earns its keep.
The time at stake is real. A 2026 study by Deming and colleagues (NBER Working Paper 35706, a survey of about 9,000 workers in Norway) found the average worker spends 4.7 hours a week in meetings, roughly 12% of the working week, and meetings account for about 14% of firms' wage bills. Microsoft's 2023 Work Trend Index put it more bluntly: the average employee spends 57% of their time in meetings, emails and chats, leaving 43% for focused work. An older study by Otter.ai and Professor Steven Rogelberg of UNC Charlotte (632 workers, summer 2022) counted 17.7 meetings a week totalling 18 hours, and estimated unnecessary attendance costs about $25,000 per employee per year. Every one of those meetings generates notes somebody has to write. That is the chore AI is actually good at.
Three ways to get AI meeting notes, and which one fits your week
There are three routes, and they suit different kinds of work. Pick the one that matches how your meetings actually happen, not the one with the flashiest demo.
| Route | How it works | Best for | The catch |
|---|---|---|---|
| Paste-in chatbot (ChatGPT or Claude) | You paste a transcript or your rough notes into the chat and ask for notes and action items. | Occasional meetings, sensitive calls, people who want full control. | You need a transcript first, from the meeting app or a recorder. |
| Auto-recorder (Otter.ai or Fireflies.ai) | A bot joins the call, records, transcribes, and produces notes and action items by itself. | Back-to-back meetings, teams that want notes without thinking about it. | Needs a paid plan for serious use, and everyone in the call must know it is recording. |
| Built-in notes (Zoom, Teams, Google Meet) | The meeting app's own AI summary, turned on in the call settings. | Teams already living inside one platform. | Summaries are shorter and less customisable, and quality varies by platform. |
If you attend two or three meetings a week, the paste-in route is the right start. It costs nothing, it works with any meeting, and you stay in charge of what goes in. If you live in meetings all day, an auto-recorder pays for itself in the first week. Start cheap, and only pay when the free route is clearly costing you time.
Start with a transcript, not a blank prompt
The single biggest mistake is asking AI to write notes from memory, or from three lines you scribbled on a pad. AI needs the raw material. Get a transcript first. Zoom, Teams and Google Meet can all produce one from the call settings. If your meeting was in person, record it on your phone (with everyone's agreement, more on that below) and run it through a transcription tool. Our guide to the best AI transcription tools covers the free options that handle this step.
A rough transcript is fine. It does not need to be perfect. What it needs is the actual words people used, because the words are what stop the AI inventing decisions. A transcript that says "Sarah: I think we could maybe ship on Friday if QA is free" is honest material. A prompt that says "write notes about the product launch meeting" with no transcript is an invitation to invent.
The three prompts that actually work
One prompt is not enough. A meeting produces three different things, and each needs its own instruction. Copy these, paste your transcript underneath, and change the bracketed parts.
Prompt 1: the notes.
"Turn this meeting transcript into notes. Structure: (1) Decisions made, one line each, with who made the decision. (2) Key points discussed, short bullets. (3) Open questions, things nobody decided. Rules: only include decisions that are stated clearly in the transcript. If something was suggested but not agreed, put it under open questions, never under decisions. Keep the whole thing under 300 words. Transcript: [paste it here]"
Prompt 2: the action items.
"From this meeting transcript, extract every action item into a table with four columns: Task, Owner, Due date, Notes. Rules: only list a task if a specific person was named as the owner in the transcript. If no owner was named, write 'Owner not assigned'. Only include a due date if one was actually stated; if none was stated, write 'No date given'. Do not guess owners or dates. Transcript: [paste it here]"
Prompt 3: the follow-up email.
"Write a short follow-up email based on these meeting notes. Keep it under 150 words. Start with the one decision that matters most. List the action items as bullets with owner and date. End with the date of the next meeting if one was mentioned. Plain, direct language, no corporate jargon. Notes: [paste the notes from prompt 1]"
The rules inside prompts 1 and 2 do the real work. "Only include decisions that are stated clearly" and "do not guess owners or dates" sound obvious, but without them the model treats a suggestion as a decision and a passing mention as a deadline. With them, the output stays inside what was actually said, and the gaps show up as gaps instead of as fiction.
A worked example: from a messy transcript to clean action items
Here is a short sample transcript I made up for this article. It is the kind of six-line exchange every office has had this week:
"Right, so the pricing page. Tom: I can get the draft to Sarah by Thursday, I think. Sarah: Thursday works if marketing signs off first. Mark: and the launch date, are we still saying the 20th? Tom: the 20th is tight but doable, if QA clears it. Sarah: I will chase QA today. Mark: okay, and someone needs to tell the client the date might move. Tom: I can do that after QA replies."
Run prompt 2 over that transcript and the action items come back like this:
| Task | Owner | Due date | Notes |
|---|---|---|---|
| Send pricing page draft to Sarah | Tom | Thursday | Needs marketing sign-off first |
| Chase QA on the launch date | Sarah | Today | Launch date is 20th, tight but doable if QA clears it |
| Tell the client the date might move | Tom | No date given | Only after QA replies |
Read that table against the transcript and notice two things. First, the launch date itself is not an action item. Nobody was asked to set it, so it does not appear as a task, and that is correct. Second, "tell the client" has no date, because Tom never gave one. A careless prompt would have written "Thursday" next to it. The rule "do not guess dates" is what kept that cell honest, and it is the difference between notes people trust and notes people stop reading.
Now compare with what happens when you skip the rules. The same transcript, a one-line prompt ("turn this into action items"), and the table comes back with Tom owning the launch date, Sarah owning the client call, and a due date of Thursday on everything. It looks tidy. It is also wrong in three places, and the people named would find out the hard way.
What AI gets wrong, and the two-minute check that catches it
AI meeting notes fail in the same four ways, and once you know the pattern you can spot all four in about two minutes.
- A suggestion becomes a decision. "We could ship on Friday" turns into "Decision: ship on Friday." This is the most common error, and the one that causes real damage. Check every line under "Decisions" against the transcript. If the words "maybe", "could" or "I think" were in the room, it was not a decision.
- An owner gets invented. The AI assigns a task to the person who spoke most, or to the meeting organiser, when nobody actually agreed to do it. Check every owner against the transcript. If no name was attached, the owner is "not assigned", and saying so is better than naming the wrong person.
- A date gets borrowed. One date in the meeting (say, Thursday for the pricing draft) migrates onto every other task. Check each date separately. Dates do not transfer between tasks.
- Names get mixed up. Similar names, misheard words in a transcript, and two people with the same first name all cause this. If your transcript has a "Sam" and a "Sarah", read those lines twice.
The check itself is simple. Read the notes once, pen in hand, and ask three questions: did we actually decide this, did this person actually agree to it, and was this date actually said? Anything that fails a question goes back to the transcript. Two minutes, every meeting, and the notes stay trustworthy. Skip the check and the notes become a liability, because a wrong action item with your name on it is worse than no notes at all.
Do you need a paid plan?
Probably not yet. The free tiers cover most people's meeting load, and the paid plans buy convenience, not better notes.
ChatGPT's free plan handles prompts 1 to 3 with a pasted transcript, and ChatGPT Plus costs $20 a month (OpenAI's pricing page, checked 9 November 2026). Plus buys more messages and file uploads, not smarter summaries. Claude's free plan does the same job, and Claude Pro is $20 a month billed monthly, or $17 a month if you pay annually (Anthropic's pricing page, checked 9 November 2026). For paste-in notes, either free plan is enough, and the choice comes down to which chatbot you already use. The same free-tier logic applies to the other writing jobs we cover on this blog: if you can write a cover letter with ChatGPT on the free plan, you can write meeting notes on it too.
Auto-recorders are where the money starts. Otter.ai's free Basic plan gives 300 transcription minutes a month, with a 30-minute limit per meeting, and Pro is $16.99 a month billed monthly or $8.33 a month billed annually (otter.ai/pricing, checked 9 November 2026). Fireflies.ai's free plan is $0 with limited storage, and Pro is $10 a seat a month billed annually or $18 billed monthly (fireflies.ai/pricing, checked 9 November 2026). If you sit in more than five hours of meetings a month, the free Otter tier runs out fast, and that is the honest point to pay. Below that, free is fine.
One rule before you press record
Tell people before you record a meeting. It is a courtesy, and in many places it is also the law. The United States has a federal one-party consent baseline, which means one person in the conversation can consent to recording, but several states require every party to consent. Other countries have their own rules, and some employers have their own policies on top. This article is not legal advice. The practical rule is simple: say at the start of the call that you are recording for notes, and give anyone who objects a chance to say so. The auto-recorder bots announce themselves when they join, which helps, but it is still your job to make sure nobody is surprised.
The same courtesy applies to sharing. Notes from a meeting belong to the people who were in it. Do not forward them to someone who was not there without checking with the group first, especially when the notes contain something sensitive.
Frequently asked questions
Can I trust AI meeting notes without checking them?
No. AI turns suggestions into decisions, invents owners and borrows dates, and the mistakes look confident. The notes are a draft. Read them once against the transcript, ask whether each decision was really made and each owner really agreed, and fix what fails. Two minutes is enough.
Is it OK to record a meeting for AI notes?
It depends on where you are and who is in the room. Tell everyone at the start of the call that you are recording and why, and check your local rules before you do it regularly. Several US states require all-party consent, and other countries have their own rules. When in doubt, ask first.
Do I need a paid plan to use AI for meeting notes?
No. ChatGPT and Claude both have free plans that handle the prompts in this guide, and Otter.ai's free plan covers 300 transcription minutes a month. Pay only when the free limits run out, which for most people means more than five hours of meetings a month.
What is the best AI tool for meeting notes?
It depends on how many meetings you attend. For two or three meetings a week, paste a transcript into ChatGPT or Claude: free, and you stay in control. For all-day meetings, Otter.ai or Fireflies.ai record and summarise automatically. For teams already inside Zoom, Teams or Google Meet, the built-in summary is the simplest option, though it is less customisable.
How long should AI meeting notes be?
One page. Decisions in one-line bullets, action items in a table with owner and date, open questions at the end. If the notes run longer than a page, the meeting probably had too many decisions for one session, and that is worth knowing too.
The habit that makes it work
AI meeting notes are not really about AI. They are about a habit: transcript in, rules on, two-minute check, notes out. The tool does the typing. You do the judging. Get that split right and you get your twenty minutes back after every meeting, and your colleagues get notes they can actually rely on. Get it backwards, trust the draft without reading it, and you will spend far more than twenty minutes fixing the confusion it caused.
Start with one meeting this week. Record it, paste the transcript into ChatGPT or Claude with prompts 1 and 2, run the two-minute check, and send the result. If the notes come back clean, you have a workflow. If they come back wrong in the ways described above, you now know exactly which rule to tighten. Either way, you will know more about your meetings than you did before, and that is the point. And when a decision from the meeting needs chasing, our guide on how to use AI to write a follow-up email after no reply shows how to turn those action items into a nudge that actually gets an answer.
By Kashan Ajmeri | Last updated: 9 November 2026
Sources: Deming, Loken, Willen and Xu (2026), NBER Working Paper 35706, "Meetings", via CEPR VoxEU; Microsoft 2023 Work Trend Index Annual Report; Otter.ai and Professor Steven Rogelberg (UNC Charlotte) meeting study, summer 2022; OpenAI Help Center, "What is ChatGPT Plus?" (checked 9 November 2026); Anthropic pricing page (checked 9 November 2026); otter.ai/pricing (checked 9 November 2026); fireflies.ai/pricing (checked 9 November 2026); DEV Community, "How to Use AI to Turn Meeting Notes Into Action Items" (October 2026).
Related on AI Tools Hub: Best AI Transcription Tools for Meetings and Interviews | Best AI Note-Taking Tools for Real Work | Can AI Help Me Write a Complaint Letter?
