Meeting recording to searchable knowledge base with transcripts, summaries and action items

Your team just finished a crucial strategy meeting. It was rich with decisions, nuances, and commitments. Three weeks later, someone asks: "Wait, who was supposed to handle the customer integration redesign?" Nobody remembers. The recording exists somewhere, but searching through 90 minutes of video is impractical, and the notes from that day are scattered across Slack, email, and someone's notebook. This scenario repeats constantly in distributed teams. Meetings generate enormous amounts of value in the moment, then evaporate because the information gets trapped in unsearchable recordings and fragmented notes. Action items slip. Context gets lost when new team members join. Decisions get remade because nobody can cite what was actually agreed. The solution is not to record fewer meetings or take better notes. It is to build a system that automatically extracts, structures, and surfaces the information from every recording. MeetGeek handles transcription and summarisation, NextStep tracks the resulting action items, and Shownotes AI provides a second layer of transcription quality and multilingual support. Connect these with an orchestration platform, and you have a knowledge base that stays current without manual intervention.

The Automated Workflow

This workflow captures meetings in MeetGeek, extracts structured data, summarises outcomes, and routes action items into a tracking system via NextStep. Zapier is the orchestration tool here because it integrates cleanly with all three platforms and requires minimal setup. Start by enabling MeetGeek's native webhook integration. When a meeting ends, MeetGeek fires a webhook containing the meeting ID, transcript, and initial summary. You configure Zapier to catch that webhook.

CopyPOST https://hooks.zapier.com/hooks/catch/YOUR_ZAPIER_WEBHOOK_ID/

MeetGeek sends a payload that looks like this:

Copyjson
{ "meeting_id": "meet_abc123", "title": "Q1 Planning Session", "duration_minutes": 87, "transcript": "full transcript text here", "summary": "initial AI summary", "attendees": ["alice@company.com", "bob@company.com"], "recording_url": "https://meetgeek.com/recordings/abc123"
}

In Zapier, your first action extracts key action items and decisions from the transcript. You do this by passing the transcript through Claude Sonnet 4.6 via the Anthropic integration. Zapier's formatter lets you construct a prompt:

CopyYou are an action item extractor. Read the following transcript and identify:
1. All explicit action items with assigned owners
2. Key decisions made
3. Important dates or deadlines mentioned Format as JSON with keys: action_items (array), decisions (array), deadlines (array). Transcript:
{{transcript}}

Claude returns structured JSON. Zapier extracts the action items array and loops through each item, creating a new task in NextStep for every one.

CopyPOST https://api.nextstep.io/v1/workflows
Authorization: Bearer YOUR_NEXTSTEP_API_KEY
Content-Type: application/json

For each action item, you send:

Copyjson
{ "title": "Redesign customer integration flow", "owner_email": "alice@company.com", "due_date": "2026-04-15", "description": "From Q1 Planning: Review current flow and propose new UX patterns", "status": "pending", "metadata": { "source": "meeting", "meeting_id": "meet_abc123", "recording_url": "https://meetgeek.com/recordings/abc123" }
}

NextStep creates the workflow task and assigns it to Alice with automatic reminders.

Adding Shownotes AI for redundancy.

If you want a second transcription or better multilingual handling, you can route the recording URL from MeetGeek to Shownotes AI in parallel. Shownotes AI re-transcribes the audio and produces its own summary. You then run a comparison in Claude to identify any critical details that MeetGeek missed.

CopyPOST https://api.shownotes.ai/transcribe
Authorization: Bearer YOUR_SHOWNOTES_API_KEY

Copyjson
{ "recording_url": "https://meetgeek.com/recordings/abc123", "language": "en", "output_format": "detailed"
}

Shownotes AI returns its transcript. You feed both transcripts into Claude with this prompt:

CopyCompare these two meeting transcripts and identify any action items, decisions, or dates that appear in one but not the other: Transcript A (MeetGeek):
{{meetgeek_transcript}} Transcript B (Shownotes):
{{shownotes_transcript}} Return only the items that are missing or significantly different.

Any gaps get flagged, and you can manually review them or create additional NextStep tasks if the secondary transcript caught something important.

Final step: structured storage.

Store the meeting record in a searchable system. Most teams use Notion, Airtable, or a custom database. You can add a Zapier step that creates a Notion page or Airtable record containing the title, date, attendees, summary, key decisions, action items, and a link to the recording.

CopyPOST https://api.notion.com/v1/pages
Authorization: Bearer YOUR_NOTION_TOKEN

This ensures that when someone searches for "customer integration" three months later, they find the meeting, the full transcript, and the original action items all in one place.

The Manual Alternative

If you prefer to retain full control over extraction and prioritisation, you can skip the automation and use these tools as a semi-manual workflow. After a meeting, download the recording from MeetGeek. Open Shownotes AI, upload the file, and let it generate the transcript and summary. Copy both into a shared document. Manually read the transcript, highlight action items, and type them into NextStep yourself. This takes 10-15 minutes per meeting but gives you the chance to contextualise each item and adjust owners or deadlines before they enter your tracking system. The downside is obvious: this approach does not scale. By meeting four or five, you are spending an hour per week on extraction alone. Automation solves this by doing that work while you focus on executing the action items, not documenting them.

Pro Tips

1. Handle transcript errors gracefully.

Neither MeetGeek nor Shownotes AI produces perfect transcripts, especially if your meetings include heavy accents, background noise, or technical jargon. Always route unclear action items through a human reviewer. Add a "needs review" flag to any task where Claude had low confidence (below 80%) when extracting the owner or deadline.

2. Rate limits and costs.

MeetGeek's webhook fires immediately after a meeting ends. If you have many meetings, your Zapier zaps will run frequently. Zapier's free tier allows 100 tasks per month; move to a paid plan if you exceed this. Claude Sonnet 4.6 is reasonably priced for transcript analysis. At roughly £0.003 per 1000 tokens for input, analysing an 8000-token transcript costs under £0.03. Shownotes AI charges per minute of audio transcribed, so run it only for meetings where you suspect MeetGeek may have missed something.

3. De-duplicate action items.

If the same task gets mentioned twice in a meeting, Claude may extract it twice. Build a Zapier step that checks NextStep for existing tasks with similar titles before creating a new one. This prevents duplicate reminders and conflicting due dates.

4. Store metadata wisely.

Always include the original meeting ID, recording URL, and transcript snippet in your NextStep metadata. This lets Alice click through to the exact moment in the recording where her action item was assigned, rather than relying on memory or email.

5. Version control summaries.

If you re-analyse a transcript (because you added Shownotes AI after the fact), keep both summaries. Store them in Notion or Airtable with a "summary_version" field. This helps if you need to audit why an action item was created or missed.

Cost Breakdown