Can AI extract action items from meeting notes?

Yes, reliably, if you tell it what an action item is, which fields you want, and to use only the notes. Ask for a table or JSON so the result pastes straight into a to-do list or spreadsheet.

Why ask for JSON instead of a list?

JSON has named fields, so every item has an owner, a task and a deadline in the same place. It pastes into spreadsheets, task apps and scripts without retyping. For a human-only read, a table does the same job.

How do I check an extraction is complete?

Read the source once, top to bottom, ticking each item as you meet it in the extraction. Anything unticked in the source was missed; anything in the extraction you did not tick was added. One pass, a minute for a page of notes.

Extract information and action items — AI With PDFs

Prose answers make you the last step in your own process — you still have to read them and type the tasks somewhere. Ask for a structure and that step disappears. A table pastes into a spreadsheet; JSON pastes into almost anything. The model is good at this because extraction is finding, not inventing.

The extraction prompt, piece by piece

1
Say what to find
"Every action item" is clearer than "the important bits". Define it: a task someone agreed to do.
2
Say the fields
Owner, task, deadline. If a field is missing in the source, tell the model to write "not stated" rather than guess.
3
Say the format
A table with named columns, or JSON with named keys. Name them and the model will use them.
4
Forbid additions
"Use only what the notes say." Extraction goes wrong when the model completes a task that was left open.

Try it

Action items from meeting notes, as JSON

From the meeting notes below, extract every action item. Return only valid JSON: an array of objects with keys "owner", "task", "deadline". Use the deadline exactly as written in the notes; write "not stated" if there is none. Do not include decisions that were deferred. Use only what the notes say.

Notes:
Hall committee meeting — 9 September 2026, 7 pm. Present: Tom (chair), Aisha (treasurer), Ben, Priya. Apologies: Sam.

Tom opened with the caretaker's report: two lost keys in August, both charged. Aisha said the hire income for August was $2,140, up from $1,870 in July, mostly from three Saturday bookings. Ben raised the broken heater in the committee room again — it has been out since June. Tom agreed Ben would get two quotes for a replacement before the next meeting. Priya reported that the new website booking form is live but the confirmation email is going to spam for some users; she will speak to the web provider this week and report back by Friday 19 September. Aisha will move the deposit to $250 in the terms document and circulate the revised version to the committee by 16 September for comment. Tom will write to the three regular hirers about the rate change before 30 September. The next meeting is Tuesday 14 October at 7 pm. Ben asked whether the hall should offer a discount for weekday daytime bookings; no decision was made and it was carried over to the next agenda.

Three free runs a day on the site key. Add your own key for unlimited runs.

Did it work? Valid JSON with four items. Now do the one-pass check: read the notes once, top to bottom, and tick each action as you meet it. Ben's quotes, Priya's email fix, Aisha's terms, Tom's letter. Four ticks, done.
The trap: the weekday discount. It was *discussed* and *carried over* — nobody agreed to do anything. If it appears as an action item, the model turned a topic into a task. The instruction *"do not include decisions that were deferred"* is there to stop exactly that.

What a correct extraction looks like. Compare yours line by line.
Owner	Task	Deadline
Ben	Get two quotes for a replacement committee-room heater	Before the next meeting (14 October)
Priya	Speak to the web provider about confirmation emails going to spam	Friday 19 September
Aisha	Move the deposit to $250 in the terms document and circulate it	16 September
Tom	Write to the three regular hirers about the rate change	Before 30 September

Knowledge check

The extraction includes: {"owner": "Ben", "task": "Set up a weekday discount", "deadline": "not stated"}. What went wrong?

Extract information and action items

Lesson FAQ

Stop overpaying for AI tools