Never Write Meeting Notes Again: AI Automation for Meeting Documentation

Short answer: A meeting-minutes workflow takes your audio recording, transcribes it with Whisper, and has Claude turn the transcript into a structured summary with decisions and action items. It then emails everyone automatically. It replaces the note-writing, formatting and distribution that follows a meeting, which is where the time goes.

What this is: the workflow, written up so you can judge the engineering, with the n8n JSON to download and import. The widget on the demos page does not currently transcribe: it needs a transcription service that is not deployed on the machine hosting the demos, so it returns a sample summary instead. The other demos are live. I would rather tell you that than have you find out by uploading a file.

A 1-hour meeting generates 20-30 minutes of follow-up work: writing the notes, formatting them, sending them to all attendees, logging action items in the project manager. Who does all that? Usually the person who already has too much on their plate.

And what happens 48 hours later? Nobody read the notes. Action items sit untouched. In the next meeting, everything repeats.

The problem isn’t the meeting itself, it’s the manual documentation afterward.

What Meeting Documentation Actually Costs

Consider this: 10 meetings per week, 20 minutes of follow-up each. That’s 200 minutes, over 3 working hours, lost weekly to meeting notes. Over a year that is around 170 hours, or about four working weeks spent on documentation alone.

And that’s before accounting for quality problems:

• Notes are incomplete because you can’t listen and write simultaneously
• Action items get forgotten or phrased unclearly
• Decisions are buried in prose, not findable
• Different attendees remember things differently

How does the meeting minutes bot work?

I built an n8n workflow that turns an audio recording into a structured meeting summary, completely automated.

Step 1: Transcription with Whisper

OpenAI’s Whisper model is purpose-built for speech and handles the things that break naive speech-to-text:

• Accents and dialects
• Technical jargon and company names
• Multi-speaker conversations
• Poor audio quality (video call background noise)

How long a transcription takes depends on the length of the recording and on whether you call a hosted API or run the model yourself, so measure it on your own audio rather than trusting a number from someone else’s setup. The local options below are meaningfully slower than the hosted one, which is the trade you make for keeping the audio in-house.

Step 2: Structuring with Claude

The raw transcript goes to Claude Sonnet, which extracts a structured protocol:

# Meeting: Q2 Planning: March 10, 2025

**Attendees:** Anna, Stefan, Maria

## Summary
- Q2 budget increased 15%, focus on marketing
- New CRM rollout moved to July
- Customer service team getting 2 new hires

## Decisions
1. CRM migration moved to July 1st (Owner: Stefan)
2. Budget increase approved (Owner: Anna)

## Action Items
- [ ] Stefan: Issue RFP for CRM vendors by March 20
- [ ] Maria: Post job listings by March 15
- [ ] Anna: Update budget in planning doc by March 12

## Open Questions
- Will we need external consultants for CRM migration?

## Next Meeting
March 24, 2025 at 2:00 PM

Claude uses structured output (Tool Use) so the protocol always follows the same format, regardless of how chaotic the discussion was.

Step 3: Distribution

The finished summary is automatically emailed to all attendees. Optionally: archived to Google Docs, Confluence, or Notion.

End to end, the run is dominated by the transcription step. Everything after it is one model call and an email.

Privacy: What You Need to Know (GDPR)

Meeting recordings are sensitive. Before implementing, three important points:

1. Consent is mandatory

You may only record meetings when all participants have consented. Easiest approach: standard language at the meeting start, confirmed by a checkbox in the calendar invite.

2. Cloud vs. local processing

The workflow as downloaded sends audio to OpenAI’s Whisper API. For companies with strict data requirements, the same node can point at a local model instead:

Option	Where the audio goes	Speed	Cost
OpenAI Whisper API	Leaves your infrastructure	Fastest	Per minute of audio, billed by the provider
whisper.cpp (local)	Stays on-premise	Slowest, CPU-bound	Server costs only
faster-whisper (local)	Stays on-premise	Between the two, faster with a GPU	Server costs only

Check the provider’s current per-minute rate rather than a figure in a blog post, and benchmark the local options on your own hardware. Both change.

3. Retention is something you configure, not something you get

The workflow does not write the audio to disk, but that is not the same as deleting it. n8n keeps execution data, including binary payloads, according to its own pruning settings, so an audio file can sit in the execution history long after the summary has gone out. Set the retention policy on your n8n instance deliberately, and treat “the workflow doesn’t store it” as insufficient for a GDPR record.

ROI: do the arithmetic with your own numbers

The saving is whatever your follow-up work actually costs, so the only useful version of this table is the one you fill in:

Scenario	Without Automation	With Bot
Writing notes	~20 min/meeting	0 min
Logging action items	~10 min/meeting	0 min
Email to attendees	~5 min/meeting	0 min
Total	~35 min/meeting	~2 min (upload)

Worked through: 10 meetings a week at 33 minutes saved is about 5.5 hours a week, near 24 hours a month. At €50 an hour that is roughly €1,200 a month in time, against a bill in the low tens of euros for the API calls.

Two honest caveats. The minutes still need reading, and someone has to chase the action items, so treat this as time moved rather than time erased. And if your meetings currently produce no notes at all, automation does not save you 35 minutes: it gives you documentation you did not have, which may be worth more, but it is not the same claim.

Use Cases

The bot works especially well for:

Weekly team standups Short, action-focused notes. The top 3 tasks from the 15-minute meeting land automatically with the responsible people.

Client calls Formal summaries with clear next steps. The client receives the notes while you’re still clicking “leave meeting.”

Strategy and planning sessions Long meetings with many decisions. Claude structures even complex discussions clearly.

1:1 meetings Development conversations, feedback sessions. Notes are stored privately, not distributed to everyone.

What You Should NOT Automate

Sensitive HR conversations: disciplinary meetings, terminations, salary negotiations. Human judgment and discretion matter more than efficiency here.

Confidential contract negotiations: when external parties are involved without recording consent.

The system doesn’t automatically detect “off-limits” topics, that responsibility stays with you.

Technical Stack: n8n + Whisper + Claude

The workflow runs on n8n (self-hosted or cloud) and uses:

• OpenAI Whisper API for transcription
• Claude Sonnet (Anthropic) for structured extraction
• SMTP for email delivery
• Optional: Google Docs API for archiving

At 10 hours of meeting audio a week, the running cost is two line items and a server you are probably already paying for:

Item	How it is billed
Whisper API	Per minute of audio: 40 hours a month is 2,400 minutes, times the provider’s current rate
Claude API	Per summary, driven by transcript length: a 60-minute meeting is a few thousand tokens in, a few hundred out
n8n (self-hosted)	€0 beyond the server

I have deliberately not put euro figures in that table. Both providers have changed their prices more than once since this workflow was built, and a stale number that flatters the case is worse than no number.

Download the Workflow

Not a screenshot: the real workflow. This is the n8n JSON: import it and inspect every node yourself.

→ Meeting Minutes Bot Workflow (JSON)

It is the portable version, wired to OpenAI Whisper and Claude, so it works on your instance with your keys. My own instance runs a variant of it pointed at different providers. Where they differ is documented in the technical teardown.

Requirements:

• n8n (self-hosted or cloud)
• OpenAI API key (for Whisper transcription)
• Anthropic Claude API key
• SMTP configuration for email delivery

Setup in 3 steps:

1. Import JSON via n8n → Settings → Import Workflow
2. Add API keys to credentials
3. Note the webhook URL and connect it to your upload interface

Technical detail

If you’re interested in the details behind the workflow: why a forced schema beats asking for prose, the difference between tool-use and JSON mode, and which node does the shaping:

→ Building an AI Meeting Assistant with Whisper, Claude, and n8n (leinss.xyz)

Your Next Step

How many meetings do you have per week, and how long does the follow-up actually take? Multiply the two. That is what you are deciding about.

If it is more than a couple of hours a month, this is worth building.

This sits alongside my other communication automation work, and you can see related workflows on the live demos page.

---

Want to set this up for your team or adapt it to your specific tools (Confluence, Notion, Teams)? Book a free discovery call.