You can automate meeting notes with AI by feeding an audio recording into a transcription service, passing the raw text to a Large Language Model (LLM) that extracts a concise summary and a list of action items, then routing those items into Notion for reference and Asana for execution - all orchestrated in n8n. The result is a hands-free workflow that turns every meeting into a searchable knowledge base and a ready-to-act task list without manual copy-pasting.
| Tool | Plan / Price | Role |
|---|---|---|
| Fireflies.ai | Free tier (30 min transcription/month) - paid plans start at $10 / mo | |
| Record meeting, generate raw transcript | ||
| OpenAI GPT-4 (Chat Completion) | Pay-as-you-go: $0.03 / 1 K prompt tokens, $0.06 / 1 K completion tokens | Summarize transcript, extract action items |
| n8n (self-hosted Docker) | ||
| Free (Community Edition) - n8n.cloud starts at $20 / mo for 2 M executions | ||
| Glue everything together, schedule, conditional routing | ||
| Notion | Free tier (up to 1 k blocks) - Personal Pro $5 / mo | Store meeting minutes and summary |
| Asana | Free tier (up to 15 members) - Premium $13.99 / mo per member | Create assigned tasks from extracted action items |
| Slack (optional) | Free tier | Send a quick "meeting report" alert to the team |
Estimated build time: 2-3 hours for a first-pass workflow, plus 30 minutes for testing and tweaking.
ff.ai
). After the call ends, Fireflies will email you a link to the raw transcript (plain-text).
docker run -d --name n8n \
-p 5678:5678 \
-e N8N_BASIC_AUTH_ACTIVE=true \
-e N8N_BASIC_AUTH_USER=admin \
-e N8N_BASIC_AUTH_PASSWORD=YOUR_PASSWORD \
n8nio/n8n
POST
, and copy the generated URL (e.g., https://your-host.com/webhook/meeting-notes
). POST
, https://api.openai.com/v1/chat/completions
. In Authentication, choose Header Auth and add Authorization: Bearer {{ $env.OPENAI_API_KEY }}
(store the key in n8n's Credentials → API Key).
Use the following JSON body (this is the prompt that extracts both a summary and bullet-point actions):
{
"model": "gpt-4o-mini",
"temperature": 0,
"messages": [
{
"role": "system",
"content": "You are an assistant that turns meeting transcripts into a short summary and a list of actionable items. Return JSON with two keys: summary (max 3 sentences) and actions (array of objects with title and assignee if mentioned)."
},
{
"role": "user",
"content": "{{ $json.body.transcript }}"
}
]
}
What this does: Sends the raw transcript to OpenAI, asking the model to output a deterministic JSON object containing a concise meeting summary and any identified action items.
{{ $json.choices[0].message.content }}
into two fields: summary
and actions
. Use an expression like {{$json["summary"]}}
after you JSON.parse
the string.Meeting - {{ $json.meetingDate }}
(extract date from webhook payload) {{$json.summary}}
{{ $json.actions | json }}
(store actions as raw JSON for later reference) actions
array using an {{ $json.title }}
{{ $json.assignee || "" }}
(if the model identified a name, you may need a lookup table to map to Asana user IDs) {{ $json.summary }}
(provides context) #meeting-recap
with a formatted block:
*Meeting recap:* {{ $json.summary }}
*Action items:* {{ $json.actions | json }}
Using GPT-4's 8,192-token context window, a 30-minute transcript (~9 k words ≈ 13 k tokens) fits comfortably, guaranteeing full-text analysis without truncation.
| Failure mode | Why it happens | Mitigation |
|---|---|---|
| Transcript inaccuracies | ||
| Fireflies' speech-to-text can mis-recognize jargon or overlapping speakers. | Record in a quiet environment, enable "high-quality transcription" (paid tier). Add a small n8n Function node that runs a spell-check on the transcript before sending to OpenAI. | |
| OpenAI token limits | ||
| GPT-4's context window is capped at 8,192 tokens. Very long meetings (> 45 min) may exceed it. | Summarize the transcript in chunks (split at speaker changes) and feed sequentially, then combine the partial summaries. | |
| Rate-limit / quota | ||
| OpenAI enforces a per-minute request cap (~60 rpm for pay-as-you-go). Fireflies may fire multiple webhooks quickly. | Add an n8n Delay node (e.g., 1 second) before the OpenAI call, and enable Rate Limit in n8n's settings. | |
| Auth token expiry | ||
| Fireflies webhook URLs and OpenAI API keys can be rotated. | Store keys in n8n Credentials, set a reminder to rotate every 90 days. Use n8n's Cron node to ping the webhook URL monthly to ensure it's still reachable. | |
| Cost blow-up | ||
| OpenAI usage is per-token; a 30-minute transcript can cost ~$0.78 (13 k prompt × $0.03/1 k). Repeating daily adds up. | Enable a Switch node that only runs the OpenAI step if the transcript length > 2 k tokens, otherwise skip. Monitor usage in the OpenAI dashboard, set a spending alert. | |
| Assignee mapping failures | ||
| The LLM may output free-form names that don't match Asana user IDs. | Maintain a simple CSV file in n8n (Read Binary → Parse CSV) that maps "John Doe" → 1234567890 . Use a Function node to replace assignee strings before the Asana request. |
|
| Network timeouts | ||
| Large payloads to Notion or Asana can exceed default n8n timeout (30 s). | Increase Request Timeout in the HTTP Request node (e.g., 120 000 ms) or enable Retry with exponential backoff. |
For a deeper technical reference, see n8n's documentation.
The summary is deterministic because the prompt sets temperature: 0
. In practice, GPT-4 produces a concise 2-sentence recap that matches human-written minutes about 95 % of the time for clear audio.
Yes. Any service that returns plain-text via webhook (e.g., Otter.ai, AssemblyAI) works - just point the webhook URL to the n8n trigger and map the payload field name to transcript
.
Fireflies offers a Teams integration; the webhook payload format is identical. Follow the same n8n steps, only the source of the webhook changes.
Store all secrets (OpenAI key, Asana PAT, Slack token) in n8n Credentials, not in plain code. Enable HTTPS on your n8n instance (use a reverse proxy with Let's Encrypt) and restrict the webhook URL with a secret token query param (?token=XYZ
).
Add a Function node that reads the transcript for role keywords ("owner", "designer", "PM") and maps them to Asana IDs before creating the task. This logic is pure JavaScript and lives entirely inside n8n.
Check out the Meeting-to-Action automation on our vault: https://getaab.com/vault/meeting-to-action. It includes an exportable JSON that you can import directly into n8n.
Ready to stop copying notes into Asana and Notion? Grab the free guide that walks you through every click: https://getaab.com/free. Build the workflow once, and let AI handle the rest.