Do this job with AI
Turn messy meeting notes into decisions and actions
Written by AI Tools AcademyLast checked on 27 September 2026
Helpful first: Hallucination and trust · Privacy and safety
Practice files for this page
Fernway project meeting notes (fictional) (Markdown, 1 page)
Fictional practice data. No real people or organisations.
This scenario uses a fictional set of rough notes. Download them, follow the steps in whichever AI tool you use, and compare your result with the finished version at the end.
Fernway Customer Response Project, meeting 3 notes (Markdown, 1 page)
Deliberately messy. Contains unowned actions, a moving deadline, a floated idea and an ambiguous name.
Fictional practice data. No real people or organisations.
Date: Tuesday 22 September 2026, 9:00am to 10:05am Present: Priya Shah (chair), Tom Elliott, Maya Roberts, Ravi Menon, Dan Okafor (joined 9:30), Sam Hollis (Midlands depot supervisor, on Teams for the first half hour), Leah Bennett (notes) Apologies: Sam Price and Alex Doyle (both on customer visits)
- Priya recap: project aim is every customer issue gets acknowledged and has a named owner. Update due to the board on Thurs 1 Oct.
- Website contact form: Maya confirmed it WAS going to the old sales inbox. Ravi redirected it last Weds. Found approx 60 unanswered messages in the old inbox, oldest from May.
- Maya: we can't just send a blanket sorry email to all of them
- Tom: account managers should ring the bigger accounts personally
- Agreed all 60 need reading and splitting by region before anyone replies. (who's doing the split??)
- Form auto-reply still says "we aim to respond within 24 hours". We don't. Needs rewording before more people use it.
- Ticketing system: Ravi to shortlist 3 options with rough costs. By Friday... Ravi then said actually end of next week, he's on the server migration Thurs/Fri. Tom: need something to show the board on the 1st. Priya: "see what you can do". Left there.
- Tom suggested dropping the complaints phone line and going email only, easier to track. Maya strongly against, some customers won't use email. Priya: let's think about it, park for now.
- DECIDED: from Thurs 1 Oct every complaint gets a reference number and goes in a shared tracker, even before any new system. Maya will set up the tracker spreadsheet by Weds 30 Sept.
- Budget: Dan (joined late) thinks anything over about £4k a year needs Erin's sign-off. Priya: "so let's say £4k cap then". Dan: would need to check the figure with Erin first.
- Sam said the delivery complaint numbers for last quarter are in the depot log. Sam to send them round before the next meeting.
- Someone should let the account managers know what to say if customers ask why nobody replied to their form messages.
- Priya to send the board update draft to Tom and Dan by Tues 29th for comments.
- Next meeting in two weeks, Priya to send invite.
Notes typed up quickly by Leah. Not checked with attendees.
Step 1: Understand the task
Priya wants three things: what was decided, who is doing what by when, and anything still hanging. The action list is a record people will be held to, so an owner written next to a task is a small commitment made on someone's behalf.
That changes what "good" looks like. A tidy list that fills every box is worse than a slightly untidy one that is true. Your real job is to separate what was agreed from what was said, and to make the gaps visible so Priya can close them.
Step 2: Decide whether AI is appropriate
Reasonable fit. Sorting a page of fragments into decisions, actions and questions is structuring work, and AI is quick at it. The risk is also well known: language models are built to produce complete-looking answers, so they tend to fill an empty owner or deadline with a plausible guess.
Use AI for the first sort. Keep the judgement about who agreed to what, because only you were in the room.
Step 3: Assess the data
Real meeting notes like these contain colleagues' names, internal plans, a budget figure and a reference to customer messages. That's ordinary internal business information, but it still belongs in a tool your organisation has approved for work, signed in with your work account. Don't paste real internal notes into a personal account on a free consumer tool.
These notes don't contain any customer's details, only the fact that messages exist. If your real notes did (a customer's name, a complaint about a named person, anything about someone's health or performance), take those lines out first or leave the job to a human. Removing a name doesn't always make something anonymous: "the customer in Leamington who complained about the chair" may still identify someone.
The fictional practice file is safe to use in any tool. Your organisation's policy and approved tools still take priority.
Step 4: Prepare the input
Two minutes of preparation saves ten minutes of correction.
Add the attendee list with full names and roles if your notes don't already have it. This is what lets the AI notice there are two people called Sam. Add today's date, so "Friday" and "end of next week" can be turned into dates you then check. Leave the mess in: the half-sentences and question marks are evidence of uncertainty, and you want the AI to see them.
Step 5: Choose an approach or tool
Any general assistant (ChatGPT, Copilot Chat, Claude, Gemini) can do this from pasted notes. What matters more than the tool is the instruction: you need it to report gaps instead of filling them.
In Microsoft 365 Copilot
If your organisation licenses Copilot and the meeting was transcribed in Teams, Copilot can usually draft a recap from the transcript instead of your notes, depending on your licence and your organisation's settings. A transcript records who said what; it doesn't show what was agreed, so the same checks apply. It also can't see anything said while transcription was off.
In ChatGPT, Claude or Gemini
You paste the notes in, so the tool only knows what you give it. Check which account you're signed into before pasting real work notes.
Step 6: Write the instruction
The instruction needs to say what to produce and, more importantly, what to do when the notes don't say.
Weak
Turn these notes into an action list.
Better
Using only these notes, list the decisions actually agreed, the actions with owner and deadline, and any open questions. If an owner or deadline wasn't agreed, write 'Not agreed in meeting'. Don't turn suggestions into decisions.
Why it works: It limits the AI to the notes, gives it a named way to show a gap, and warns it about the most common mistake.
A short, clear instruction is enough here. You don't need a paragraph of role-play.
Below are my rough notes from a project meeting held on [date]. Using only these notes, produce three sections. Decisions: only things the group clearly agreed. Leave out ideas that were suggested, parked or disputed. Actions: a table with Action, Owner, Deadline and Notes. If no owner was agreed, write "Not agreed in meeting". If a deadline changed or wasn't settled, write what was said and flag it. Don't infer an owner from someone's job title. If a first name could mean more than one attendee, say so. Open questions: anything I need to confirm with the chair before sending. [paste notes, including the attendee list]
Why this works: It gives the AI permission to leave boxes empty, which is the whole point.
Step 7: Inspect the answer
Here is the kind of answer a general assistant might give to the weak instruction. It looks finished, which is exactly the problem.
Decisions
- From 1 October, every complaint gets a reference number and is logged in a shared tracker.
- The budget for the new ticketing system is capped at £4,000 a year.
- The complaints phone line will be reviewed with a view to moving to email only.
Actions
| Action | Owner | Deadline |
|---|---|---|
| Set up the shared complaints tracker | Maya Roberts | 30 September |
| Shortlist three ticketing systems | Ravi Menon | Friday 25 September |
| Split the 60 unanswered messages by region | Maya Roberts | Before 1 October |
| Reword the contact form auto-reply | Ravi Menon | 1 October |
| Send last quarter's delivery complaint figures | Sam Price | Next meeting |
| Brief account managers on what to tell customers | Tom Elliott | This week |
| Send board update draft to Tom and Dan | Priya Shah | 29 September |
| Send invite for next meeting | Priya Shah | 6 October |
Read it against the notes line by line. Your memory of the meeting isn't the source. Put a mark next to every owner and every date, then find each one in the source.
Step 8: Identify likely errors
Eight actions, every box filled. Only two rows (Maya's tracker and Priya's board draft) are fully supported by the notes.
Other things to catch:
- Decision 3 overstates a parked idea. "Let's think about it, park for now" isn't a review.
- "6 October" is the AI's arithmetic for "in two weeks". It's about the meeting date, not when the invite should go.
- Tom's suggestion that account managers ring the bigger accounts has vanished. It was never agreed, so it belongs in open questions.
Step 9: Correct the result
You can fix this with one follow-up and a few manual edits. The follow-up makes the AI show its evidence, which is quicker to check than its conclusions.
For each action and decision, quote the exact line from my notes that supports the owner and the deadline. Where no line names an owner or a date, change it to "Not agreed in meeting". Two people called Sam are listed (one attended, one sent apologies): flag which Sam is meant wherever it's unclear.
Why this works: Asking for the supporting line exposes every owner or date that has no source.
Then do the part only you can do. You were in the room: if you're confident Sam Hollis offered the depot figures, say so, but mark it as your understanding so Priya can correct you. Turn "end of next week" into a date (Friday 2 October) and check it on a calendar. Keep Tom's board request next to it, because the conflict is real and Priya needs to settle it.
Step 10: Produce the finished work
Show the finished version
Customer Response Project, meeting 3 (22 September): decisions and actions
Decision agreed
From Thursday 1 October, every complaint gets a reference number and goes in a shared tracker, even before a new system is chosen.
Actions
| Action | Owner | Deadline | Note |
|---|---|---|---|
| Set up the shared complaints tracker spreadsheet | Maya Roberts | Wed 30 Sept | |
| Send board update draft to Tom and Dan for comments | Priya Shah | Tue 29 Sept | |
| Shortlist three ticketing systems with rough costs | Ravi Menon | Not settled | Ravi said end of next week (Fri 2 Oct). Tom asked for something before the board on Thu 1 Oct. |
| Check with Erin whether £4k a year is the budget limit | Not agreed in meeting | Not agreed | Dan said he'd need to check the figure with Erin first. Confirm with Dan whether he's doing this. |
| Send last quarter's delivery complaint figures from the depot log | Sam (Hollis or Price? please confirm) | Before next meeting | I think this was Sam Hollis. |
| Read the 60 or so old form messages and split by region | Not agreed in meeting | Before anyone replies | |
| Reword the contact form auto-reply | Not agreed in meeting | Not agreed | Currently promises 24 hours. |
| Brief account managers on what to tell customers | Not agreed in meeting | Not agreed | "Someone should" |
| Send invite for next meeting (about Tue 6 Oct) | Priya Shah | Not set |
Open questions for Priya
- Who is splitting the old form messages, and who is rewording the auto-reply?
- Who briefs the account managers, and should they ring the bigger accounts personally (Tom's suggestion, not agreed)?
- Does the board need Ravi's shortlist by 1 October, or is 2 October fine?
- The £4k budget cap is not confirmed until Dan has checked with Erin.
- Dropping the complaints phone line was parked. When do we come back to it?
Step 11: Reflect: did AI actually save time?
Be honest with yourself here. The first draft took under a minute. Checking every owner and date against the notes took most of the time, and you'd have spent some of that time anyway writing it by hand.
Where AI earned its place was the structure: the table made the empty cells obvious, and the evidence follow-up surfaced the second Sam faster than rereading would. Where it cost time was cleaning up confident guesses. With the better instruction from step 6, most of those guesses never appear, so the checking is quicker.
A fair rule: for long or chaotic notes, AI plus a proper check is usually quicker than starting from blank. For a short meeting with clear owners, typing it yourself wins.
Common mistakes
- Trusting a full table. A table with every cell filled looks more finished than one with gaps. In meeting notes, full usually means guessed. Treat every owner as unconfirmed until you've found it in the source.
- Letting job titles decide owners. "Tom leads the account managers, so it's Tom's job" is a reasonable guess and still a guess. Nobody agreed.
- Converting relative dates without checking. "Friday" in notes written on a Tuesday could mean this week or next, and the AI doesn't know what day the meeting was unless you tell it.
- Keeping the first version of a moving deadline. When a date changes in discussion, the AI may keep the first one it saw, or the last, or blend them. Write down what was said and flag it.
- Hiding the gaps. A footnote saying "some owners TBC" gets ignored. Put "Not agreed in meeting" in the table, where the chair will see it.
- Sending without the chair's eye. Your notes weren't checked with attendees. A quick "does this match your memory?" to Priya before posting to the whole channel costs one message.