You can automate meeting summaries end-to-end — transcription, action item extraction, and distribution — using a combination of tools like Otter.ai, Make.com, and a large language model like GPT-4o, all without writing a single line of code. The whole pipeline takes about two hours to set up and will save most teams four to six hours per week on follow-up work. Here’s exactly how to do it.
Why Meeting Recordings Are a Black Hole (And What To Do About It)
Most companies are drowning in recorded meetings. Zoom alone stores billions of minutes of cloud recordings, and the overwhelming majority of them are never rewatched. Someone meant to write a summary. Nobody did. Three weeks later, someone’s asking “wait, what did we decide about the pricing page?” and six people are digging through Slack.
The fix isn’t to stop recording. It’s to build a system that automatically converts every recording into something your team can actually use: a structured summary with decisions, action items, owners, and deadlines. Here’s the step-by-step playbook.
Step 1: Choose Your Transcription Tool
Before you can summarize anything, you need clean text. Three tools dominate this space right now, and they are not interchangeable.
Otter.ai
Otter.ai is the easiest to set up for teams already on Zoom or Google Meet. It joins calls as a bot participant, transcribes in real time, and stores everything in a searchable dashboard. The Pro plan runs $16.99 per user per month. The real limitation: speaker identification is inconsistent when multiple people talk fast or over each other. For internal team meetings it’s fine. For client calls with lots of crosstalk, accuracy can drop noticeably.
Fireflies.ai
Fireflies.ai is the better pick if you care deeply about CRM integration. It connects natively with HubSpot, Salesforce, and Notion, and its Pro plan ($18/seat/month) includes AI summaries built in. The catch: the built-in summaries are generic. They pull “key points” that often miss the actual decisions. You’ll want to replace its native summarization with your own prompt, which I’ll cover in Step 3.
Whisper via AssemblyAI
If you’re dealing with recorded audio files rather than live calls — think recorded customer interviews, webinars, or async video updates — AssemblyAI gives you API access to a Whisper-class model with speaker diarization. Pricing is usage-based, around $0.65 per hour of audio, which makes it extremely cheap at scale. This is the route I’d recommend for any team processing more than 50 hours of recorded content per month. The tradeoff is that it requires a bit more technical plumbing to connect to your downstream tools.
Step 2: Get Your Transcript Into Your Automation Platform
Once transcription is handled, you need to pipe that text somewhere it can be processed. Make.com (formerly Integromat) is my preferred tool here. It’s more powerful than Zapier for multi-step logic, and the pricing is more honest — the Core plan starts at $9/month and includes enough operations for a small team running this daily.
If you’re using Otter.ai, set up a Make.com scenario triggered by a new Otter transcript being marked complete. Otter has a native Make.com module, so no API gymnastics required. If you’re on Fireflies, use their webhook trigger — go to Fireflies Settings → Webhooks → Add Webhook → select “Transcript Completed” and paste in your Make.com webhook URL. Make will catch the payload, which includes the full transcript text as a field you can reference in later steps.
For AssemblyAI, you’ll submit audio files via their API and poll for completion, then grab the transcript text from the JSON response. Make.com handles this fine with HTTP modules.
Step 3: Write a Prompt That Actually Extracts What You Need
This is where most teams go wrong. They use a vague prompt like “summarize this meeting” and get a vague summary back. You need to be surgical. Here’s a prompt structure that works reliably:
- Role: “You are a precise executive assistant summarizing business meetings.”
- Task: “Given the following transcript, extract: (1) a 3-sentence summary of the meeting’s purpose and outcome, (2) a numbered list of decisions made, (3) a numbered list of action items in the format [Owner]: [Task] by [Deadline if mentioned], (4) any open questions that were not resolved.”
- Constraint: “Do not invent information. If no deadline was mentioned, write ‘deadline TBD.’ If no decisions were made, say so explicitly.”
- Format: Specify exactly the output format you want — headers, bullet points, whatever lands cleanly in your destination tool.
Feed this prompt plus the transcript into GPT-4o via the OpenAI module in Make.com. GPT-4o handles long transcripts well — up to 128k tokens in context, which covers roughly a 90-minute meeting comfortably. Cost per summary run is typically under $0.10, even for long calls. If you’re running hundreds of meetings a day, GPT-4o Mini drops that to pennies, though it occasionally misses nuance on complex technical discussions.
Understanding how chaining prompts together can improve output quality is worth a detour — the concept is explained well in this plain-English guide to prompt chaining, and it applies directly here if you want to run a second pass to extract risks or sentiment.
Step 4: Route the Summary to the Right Places
A summary that lives only in Make.com’s execution log is useless. Add routing steps in your scenario to distribute it automatically:
- Slack: Post the structured summary to a dedicated #meeting-summaries channel, or better yet, use Make.com to look up the Slack channel associated with that meeting’s project and post it there. Slack’s API module in Make handles this with a simple lookup.
- Notion or Confluence: Create a new page in your meetings database with the summary pre-populated. Tag it with the meeting date, attendees, and project. This creates a searchable knowledge archive over time — something worth thinking about if you care about building a genuine AI memory layer for your business.
- Email: Send a formatted email to all attendees with the summary and action items. Include a one-line note at the top saying “This summary was AI-generated — reply to flag any inaccuracies.” That CYA line matters. People trust the output more when they know there’s a correction mechanism.
- CRM: If it was a sales or client call, log the summary and action items directly to the relevant contact record in HubSpot or Salesforce. This is where Fireflies’ native integrations actually shine, though you can replicate it in Make.com with HubSpot’s API module.
Step 5: Add a Human Checkpoint (Don’t Skip This)
Fully automated is not always fully reliable. GPT-4o occasionally hallucinates a deadline that was implied but never stated, or misattributes an action item to the wrong person. For high-stakes meetings — board calls, client kickoffs, legal discussions — build a 15-minute human review window before the summary gets distributed.
The cleanest way to do this: instead of immediately posting to Slack or email, have Make.com first create a Google Doc draft and send a single Slack message to the meeting organizer saying “Your AI summary is ready for review — approve here.” Use a Make.com webhook + a simple Google Form or even a Slack button to trigger the final distribution step. This keeps the automation fast without removing accountability.
This kind of maintenance mindset — building automation that doesn’t silently fail — is something I wrote about more directly in this guide to building AI workflows that don’t break every two weeks. The principles apply here too.
Step 6: Test on Real Meetings Before Going Live
Run your pipeline on five archived recordings before you flip it on for live meetings. Check: Did it catch the actual decisions? Did it miss any action items? Is the owner attribution correct? Adjust your prompt based on what’s failing. The most common issue I’ve seen is transcripts from meetings with heavy technical jargon — the model summarizes correctly but strips the specifics that make the action items meaningful. The fix is usually adding to your prompt: “Preserve technical terms and specific product names exactly as spoken.”
Also: test what happens when a meeting produces a transcript over 50,000 tokens. Make.com has a data transfer limit per operation. Split very long transcripts into chunks and summarize each, then run a final consolidation prompt. It’s an extra step but prevents silent failures on your longest calls.
What This Actually Costs to Run
For a 10-person team running 8 meetings per week: Otter.ai Pro for everyone ($170/month), Make.com Core ($9/month), OpenAI API costs around $3–5/month at typical meeting lengths. Total: roughly $185/month. If your team was spending even 30 minutes per meeting on manual follow-up, you’re recovering 4 hours of work per week. At a blended rate of $50/hour, that’s $800/month in recovered capacity. The ROI math isn’t close.
For smaller teams or tighter budgets, Fireflies’ Pro plan includes built-in AI summaries at $18/seat, which means you might skip Make.com and OpenAI API entirely — you just lose control over the summary format and routing logic. Trade-offs worth understanding if you’re also thinking clearly about AI automation ROI.
FAQ
What if my meetings are confidential and I can’t send transcripts to OpenAI?
This is a legitimate concern, especially in legal, healthcare, or financial contexts. You have two options: use OpenAI’s API with a data processing agreement in place (OpenAI’s API does not use your data to train models by default), or run a local model like Llama 3 via Ollama or Groq’s API, which keeps data off third-party servers entirely. Groq is surprisingly fast and cheap for summarization tasks.
Can this work with async video tools like Loom?
Yes. Loom has a Zapier and Make.com integration that can trigger on new video completion. However, Loom’s built-in transcription quality varies — I’d recommend downloading the audio and running it through AssemblyAI for anything where accuracy matters. The extra step is worth it.
Does this replace the need for a dedicated note-taker in meetings?
For most internal meetings, yes — at least for the mechanical part of capturing and distributing notes. What it doesn’t replace is someone in the room who can flag when a decision is unclear and push for resolution in real time. The AI summarizes what was said, not what should have been said. Keep a human moderator in high-stakes meetings; let the AI handle the paperwork afterward.
This article was produced with the assistance of AI, and its featured image was AI-generated. We review for accuracy, but please verify critical details.
