n8n + Notion: Build an AI-Powered Meeting Notes System

09 Jun 2026 — 4 min read

Every team has the same problem: meetings happen, notes get taken, and then nothing. Action items scatter across chat threads, follow-ups get missed, and the next meeting starts with "wait, what did we decide last time?" n8n combined with Notion and a language model fixes this — not by adding another manual step, but by removing the human from the loop entirely.

How the System Works

The core flow is straightforward. When a meeting ends, a raw transcript (from Fireflies, Otter, Google Meet, or a local recording) lands somewhere — usually an email, a webhook, or a folder. n8n picks it up, sends it to an AI model like Claude or GPT-4, and gets back structured output: a summary, a list of action items with owners, key decisions, and open questions. That structured output goes directly into a Notion database — one page per meeting, already formatted, already tagged.

No one copies and pastes. No one edits a template. The note exists in Notion within seconds of the meeting ending.

What You Need to Build It

Before touching n8n, have these in place:

A transcript source. Fireflies.ai and Otter.ai both support webhooks that fire when a transcript is ready. Google Meet exports to Drive. Even Zoom can push to S3. Pick what fits your stack.
An AI API key. OpenAI, Anthropic, or any model with a completion endpoint. Claude works especially well for structured extraction tasks with long transcripts.
A Notion integration token. Create an internal integration at developers.notion.com, share your target database with it, and note the database ID from the URL.
An n8n instance. Self-hosted on your VPS or n8n Cloud — both work identically for this workflow.

The n8n workflow itself has five nodes: a webhook trigger, an HTTP request to the AI API, a code node that parses the JSON response, a Notion node that creates the page, and an optional Slack or email node to notify the meeting organizer.

Writing the AI Prompt That Actually Works

The prompt is where most implementations fail. Asking the model to "summarize this meeting" produces summaries. Asking it to return a specific JSON structure produces data you can actually use in Notion.

A reliable system prompt looks like this:

Define the exact JSON schema you want: summary, action_items (array with task, owner, due_date), decisions, attendees, follow_up_questions.
Tell the model to infer owners from context — "John will handle the pricing review" should produce "owner": "John".
Instruct it to return nothing outside the JSON block. No preamble, no explanation.
Set a temperature of 0 or 0.1. Consistency matters more than creativity here.

In n8n, you pass the transcript as the user message and your structured prompt as the system message. The response comes back as a string — use a Code node with JSON.parse() to convert it before the Notion step.

Mapping Output to Notion Properties

Notion databases have typed properties: title, rich text, date, multi-select, people, relation. The Notion node in n8n exposes all of them. A clean mapping looks like this:

Title: Meeting date + first attendee name, or pull the meeting title from the transcript metadata.
Summary: Rich text property, populated from summary in the AI output.
Action Items: Either a rich text list in the page body, or individual rows in a related database if you want to track completion.
Attendees: Multi-select property. The AI extracts names; you map them to Notion tags.
Date: Date property set to the meeting timestamp from the webhook payload.
Status: Default to "Needs Review" — someone should still read the AI output before acting on it.

The page body can include the full structured breakdown using Notion's block API. n8n's Notion node supports appending blocks after page creation, so you can build the page body programmatically: a heading for action items, a bulleted list below, a divider, then decisions, then open questions.

If you want to skip the build-from-scratch process, ready-made n8n templates cover this exact use case — with the prompt, the Notion mapping, and the error handling already configured.

Error Handling and Edge Cases

Production deployments break in predictable ways. Build for these from the start:

Malformed AI output. Models occasionally return invalid JSON. Wrap your parse in a try/catch and route failures to a "Manual Review" Notion page instead of crashing the workflow.
Long transcripts. A 90-minute meeting exceeds the context window of some models. Add a Code node that splits the transcript into chunks and summarizes each, then combines the summaries before the final extraction pass.
Missing owners. Not every action item has a clear owner in the transcript. Default to the meeting organizer and flag it with a "Needs Owner" tag.
Duplicate triggers. Webhook-based systems sometimes fire twice. Store processed transcript IDs in a simple key-value store or Notion database and check before processing.

An AI-powered meeting notes system built on n8n and Notion is one of those automations that pays back its setup cost within a week. The meetings don't change. The discipline required to maintain notes manually disappears — and the institutional memory that used to live in someone's head starts living somewhere everyone can search.