Build a small workflow in n8n that reads your meeting notes, asks an AI model to return every action item as JSON with a task, an owner and a due date, splits that list into one item per task, and adds each one as a row in your task sheet or as a task in your project tool. The free lesson n8n: Turn Meeting Notes into Action Items in a Google Sheet builds exactly that, five steps in a line.
The trick that makes it work is asking for JSON, a fixed data shape, instead of a nice table. A table is for people to read. A task list needs each action item as a separate piece of data, and JSON gives you that.
What are the steps?
- Decide where notes come from. A Google Sheet with one row per meeting is the simplest start. A transcript from your meeting tool or a note-taker works too.
- Start with a manual trigger. In n8n, pick Trigger manually while you test. Later swap it for a schedule, a form or a new-row trigger.
- Read the notes. Add Google Sheets, Get row(s) in sheet, pick the notes spreadsheet and tab, and click Execute step to see the rows.
- Ask AI for the tasks as JSON. Add OpenAI, Message a model. Ask for a list called tasks, each with task, owner and due. Switch the prompt box to Expression so the meeting notes are filled in, and under Options set Output Format to JSON Object.
- Split the list. Add Split Out on the tasks list, so each task becomes its own item.
- Write the tasks. Add Append row in sheet on your Action Items sheet with Map Automatically, or use n8n's node for your project tool instead.
- Check the result, then publish. Read the tasks against the notes. Only when they are right, switch the trigger and click Publish.
What does this look like in a real lesson?
In n8n: Turn Meeting Notes into Action Items, you run the office at a made-up accounting group. The notes sheet holds two meetings. Get row(s) in sheet returns 2 items. The Message a model step, with GPT-4.1-MINI and the JSON Object output format, returns a list of tasks: 4 for the first meeting and 3 for the second. Split Out on the path to that list turns two meetings into 7 items, each with task, owner and due. Append row in sheet adds 7 rows to the Action Items sheet, whose row 1 holds the lowercase column names task, owner and due so automatic mapping lines up. A decision from the notes to skip the status meeting is correctly left out, because it is not a task.
The lesson ends unpublished on purpose: it starts from a button, and every run adds rows again, so you would get duplicates if you clicked twice.
What if I want a recap email too?
Do the recap in ChatGPT first. In Summarize a Meeting Transcript, you paste a bakery's Monday meeting transcript with a request for a summary, the decisions, an action items table with owner and due date, and open questions. Two dates come back as Not stated, you fill them in, and ChatGPT writes the follow-up email.
Meeting to Action joins the two: ChatGPT writes and you check the recap, then gives the action items as a JSON list, n8n turns it into four rows in a tracker with a blank Status column, and Gamma builds a five-slide kickoff deck. The lesson catches Gamma adding details nobody said in the meeting. For the one-off version, see meeting transcript to action items.
Which tools fit where?
- n8n runs the workflow and connects to sheets and task tools.
- ChatGPT or another OpenAI model pulls out the tasks.
- CoHear turns recorded sessions into notes and action items, which can be the input.
- Asana or another project tool can be the destination instead of a sheet.
If the tasks should land in a CRM instead, the meeting notes to CRM FAQ covers that route. For what counts as an action item, see the glossary entry on action items.
What mistakes should I avoid?
- Asking for a table instead of JSON. n8n gets one block of text and cannot split it.
- Leaving the prompt in Fixed mode. The AI gets the curly braces as plain text, not your notes.
- Letting AI guess owners and dates. Tell it to write "Not stated" when the notes do not say.
- Running it twice on the same notes. Plain Append adds the rows again.
- Publishing before you check. Read the first runs against the notes, and keep a person reviewing tasks that go to clients.
Key takeaways
- Ask the AI for action items as JSON with task, owner and due.
- Split the list so each task becomes one row or one task.
- Test with a manual trigger, then switch to an automatic one and publish.
- Have the AI write "Not stated" instead of guessing.
- Check the first runs against your notes.
Frequently asked questions
- Can I do this without n8n?
Yes, Zapier and Make can do the same steps. The key parts are the same everywhere: AI returns structured data and the tool turns each item into a task.
- Does my meeting tool already do this?
Many note-takers list action items in their summaries. Getting them into your own task list automatically usually still needs an integration or an automation.
- How do I stop duplicate tasks?
Run each meeting once, or match on a unique column with an append-or-update step. Marking a meeting row as processed also works.