Meeting Notes to CRM: voice-note your call debrief, find it structured in the pipeline sheet
An n8n workflow where you send a voice note to Telegram right after a sales call — transcribed automatically, then Claude extracts the deal stage, budget signals, objections, and agreed next step, updates the prospect's row in your CRM sheet, and schedules the follow-up reminder.
- 1
Voice note in
A
Telegram Triggeron voice messages, then aTelegramnode downloads the audio file. The habit that makes this work: debrief in the first three minutes after the call, while the details are still warm. - 2
Transcribe
An
HTTP Requestposts the audio to your transcription endpoint and returns plain text. Accuracy on a quiet voice note is excellent; even a car-noise debrief comes through usable because Claude reads for meaning, not word-perfection. - 3
Extract and match
Claude receives the transcript plus the current prospect list (names + companies from the sheet): identify the prospect, extract the deal fields as JSON, flag
match_confidence. High confidence proceeds; low confidence triggers the clarifying Telegram question instead of a guess. - 4
Update the row, set the reminder
A
Google Sheetsupdate writes the fields to the prospect's row (appending to notes history, never overwriting). The agreed next step's date feeds aSchedule-checked reminders tab, and the morning of, you get:📞 Follow up with Jane @ Acme — promised the case study. Pipeline reviews stop starting with 'wait, where did that deal end up?'
Frequently asked questions
How does the transcription work?
The Telegram voice file goes to a speech-to-text API — the workflow ships wired for OpenAI-compatible Whisper endpoints (including self-hosted) since that's the most common setup, with a note for swapping in any transcription service. A 2-minute debrief costs about a cent to transcribe.
How does it know which prospect I'm talking about?
Say the name in the note — 'call with Acme, spoke to Jane…' — and Claude matches it against your prospects sheet with fuzzy matching. Ambiguous or unknown names don't guess: the workflow replies asking 'Acme Corp or Acme GmbH?' and a new name offers to create a row. Wrong-row updates are the failure mode this design refuses.
What exactly gets extracted into the sheet?
Stage (discovery/proposal/negotiation/closed), sentiment, budget signals if mentioned, objections raised, promises you made, agreed next step with date, and the raw transcript link. The next-step date also creates the reminder — the follow-up you promised at 4:55 on Friday no longer depends on Friday-you remembering.