How to use AI to turn meeting notes into an action list
Take your rough notes or the transcript, hand them to an assistant, and ask it to pull out the commitments only: one line each, with an owner and a date. Everything that was discussed but not committed goes in two short separate lists, decisions and open questions. Then someone who was in the room reads the whole thing before it is circulated, and it goes out the same day. That is the entire method, and the two instructions that make it work are asking for commitments only and forbidding the tool from guessing an owner or a date.
Meeting notes are a good fit for this because the raw material is messy and the output is tightly structured. Pulling a clean list out of half an hour of back-and-forth is slow, dull, and easy to postpone until the details have faded. A tool does it immediately, which is the difference between a list that goes out while people still remember agreeing to things and one that never goes out at all.
What to feed it
A transcript if you have one, and rough notes if you do not. Both work. The notes do not need to be tidy, and cleaning them up first is wasted effort, because sorting mess is the job you are handing over. What does need to be right is a short list of who was in the meeting, with names spelled correctly, since a model working from an imperfect transcript will otherwise invent a plausible spelling and you will be fixing it forever.
Add last meeting's action list if there was one. It gives the tool the wording and format you already use, and it lets you ask for the carried-over items to be marked as still open rather than appearing as brand new tasks. That single addition is what stops the list resetting itself every week.
The tool class
Two routes, and the difference is mostly about whether you want a transcript. A meeting notetaker joins the call, records it, and produces both a transcript and a summary, which you then push through your own prompt to get the format you want. Or you paste your own notes into a general assistant after the meeting. The second route costs nothing extra, works for meetings that are not on a call, and avoids the question of whether a recorder should be in the room at all.
Whichever you use, decide in advance who is told the meeting is being recorded or processed, and where the recording ends up. Client conversations and anything involving staff matters deserve a clear answer before the first meeting, not after. It is a short conversation to have once and an awkward one to have late, and it is the same guardrail question that comes up with every setup covered in everyday automation.
Exhibit 1
One set of notes, three separate outputs, so the actions do not get buried in the discussion.
The prompt approach
Be specific about what counts and what the output should look like. Extract only things someone committed to doing. Write each as a single line: the task, one owner, one date. If the owner is unclear, write TBD. If the date was never stated, write TBD. Do not infer either. Then, separately, list decisions that were made, and separately again, list questions left open. Quote the phrase in the notes that each action came from.
The TBD rule is the whole game. Left to itself, a model will fill gaps helpfully, assigning the task to whoever was talking and putting a date on it because the format asked for one. Both are guesses that will read as facts once the list is circulated, and nobody downstream can tell them apart from the real ones. Forcing TBD turns every gap into a visible question, and a list with four TBDs on it is far more useful than a tidy one where four items are quietly wrong.
Asking it to quote the source phrase is the other line worth adding. It makes your check fast, because you can see at a glance what each action was derived from, and it discourages the model from producing items that sound like a reasonable summary of the discussion but that nobody actually agreed to.
The pass a person makes
Someone who was in the room reads the list. Not the person who is busiest, and not nobody. Four things to look for, and it takes a couple of minutes.
Exhibit 2
Four checks by someone who was there, then it goes out the same day.
Names first, because a task assigned to the wrong person is worse than no task. Then anything on the list that nobody actually committed to, which is where the quoted source phrases earn their place. Then the dates, since a model will sometimes convert "probably next week" into a specific day. Then the TBDs, which need a real answer from you rather than a shrug: an item that leaves the room without an owner is an item that will not happen.
Circulate it the same day. Not because speed is impressive, but because that is the window in which someone who heard a commitment differently will say so. A day later the list has already become the record, and correcting it feels like reopening something settled.
Why the format matters more than the tool
Everything above depends on already knowing what a good action list looks like: one task, one named owner, one real date, circulated fast. If you do not have that habit, a tool will just produce disorganized notes more quickly. The habit comes first, and it belongs alongside the other small routines that make up a steadier week. The tool then removes the friction that makes the habit hard to keep on a busy day, which is the whole reason these lists usually stop getting written.
That is the same shape as every other job worth handing over, from sorting the inbox to the rest of the routine work AI can take on. The tool does the sorting and the typing. A person decides what is true and what goes out. Keep those two straight and this is a small, safe improvement you can start using after the next meeting.