Automate Voicemail Transcription and CRM Logging with n8n

Every sales team has the same voicemail problem: listen, scribble, forget. The playback is slow, the notes are worse, and by the time anything reaches the CRM it's missing half the context. n8n voicemail automation breaks this loop by treating each incoming voicemail as a webhook event — transcribed by OpenAI Whisper, triaged by GPT-4o-mini, and filed in HubSpot before the rep's finished their next task.

The pipeline isn't complicated. Twilio fires a webhook the moment a recording is ready. n8n picks it up, downloads the audio, transcribes it with Whisper, extracts urgency and intent with a second AI call, then fans out to HubSpot, Google Sheets, and Telegram in parallel. The whole chain runs in under 30 seconds.

The biggest win isn't the transcript itself. It's knowing a voicemail is "high urgency, callback request" before anyone hits play — that's what actually changes how fast a rep responds.

What n8n voicemail automation handles

A complete voicemail workflow does more than capture text:

• Calls that go unanswered during off-hours or peak coverage gaps
• High-volume support lines where listening to every recording isn't practical
• Field teams who take calls but can't stop to type notes into a CRM
• Compliance use cases that need every phone interaction logged with a timestamp
• Triage workflows where urgency determines which queue a lead enters next

The VoiceNote Pipeline

Twilio Webhook → Download Audio → Whisper → GPT-4o-mini → [HubSpot Note | Sheets Log | Telegram Alert]

Five nodes handle the core chain. The fan-out at the end runs in parallel — the Sheets log doesn't hold up the Telegram alert. Twilio fires the webhook automatically on recording completion, so you're not polling.

Step-by-step: building the workflow

1. Receive the Twilio webhook

Twilio posts to your n8n Webhook node when a voicemail recording finishes. The payload includes RecordingUrl, RecordingDuration, From, and To. Set the webhook path to something like /twilio-voicemail and paste the full URL into Twilio under Phone Numbers → Active Numbers → Voice Configuration.

One thing that trips up early setups: Twilio appends .json to RecordingUrl by default. The actual audio file lives at the same URL without that extension. A Code node (v2, jsCode) stripping .json before the download step saves a lot of debugging.

2. Transcribe with OpenAI Whisper

The HTTP Request node (v4.2) handles the Whisper API call. Download the voicemail audio as binary using the cleaned recording URL, then POST it to https://api.openai.com/v1/audio/transcriptions with model: whisper-1 and response_format: json. The transcript arrives at $json.text.

Whisper handles phone-quality audio better than most expect. Accuracy on voicemails in the 30-to-90-second range runs above 95% for clear recordings. Longer voicemails can still clip if call quality is poor, but that's a Twilio encoding issue, not a Whisper limitation.

3. AI triage with GPT-4o-mini

A second HTTP Request node calls the Chat Completions API with a structured prompt requesting urgency level (low, medium, high), intent label (callback_request, complaint, inquiry, order, other), a 2-to-3 sentence summary, and a recommended next action. Ask for JSON output.

If you're using the @n8n/n8n-nodes-langchain.openAi v2.1 node, output arrives at $json.output[0].content[0].text — a string, not an object. Add a Code node after with JSON.parse($input.first().json.output[0].content[0].text). Don't skip this parse step. The HubSpot node downstream expects structured key-value pairs; without it, the note gets created with an empty body and you won't know until you're looking at a blank CRM record.

4. Create the HubSpot note

The HubSpot node's "Create Note" action accepts a contact phone number as a lookup value. Pass $json.From from the Twilio payload as the association key. The note body should combine the transcript and triage output so any rep can read the situation cold:

VOICEMAIL — {{urgency}} urgency
From: {{From}} | Duration: {{RecordingDuration}}s
Intent: {{intent}}

Summary: {{summary}}

Recommended next action: {{recommended_action}}

Full transcript:
{{transcript}}

If no contact matches on phone number, add a HubSpot Search node first. When the search returns empty, create the contact record, then attach the note. Orphaned notes floating without a contact record are hard to find later.

5. Fan out to Sheets and Telegram

Two branches run in parallel after the HubSpot step. Google Sheets appends one row per call: timestamp, caller number, duration, urgency, intent, summary, transcript. The Telegram branch fires only when urgency is high — an If node filters before the message goes out.

That filter isn't optional. Without it, every "just checking in" voicemail pings the team at 11pm.

Implementation patterns

Pattern 1: Urgency routing via Switch

GPT-4o-mini Triage → Switch(urgency)
  high   → Telegram alert + HubSpot task (4-hour due)
  medium → HubSpot note + Sheets row
  low    → Sheets row only

A Switch node after triage creates three execution paths. High urgency triggers a Telegram alert and a HubSpot follow-up task with a deadline 4 hours out. Medium goes to note plus Sheets. Low goes to Sheets only. Teams that've tried the "alert on everything" approach switch to this pattern within a week — quieter, and the signal is cleaner.

Pattern 2: Daily digest instead of real-time pings

Schedule (8am) → Read Sheets (yesterday) → Code (group by urgency) → Slack digest

A separate workflow runs each morning and reads yesterday's voicemail rows from Google Sheets, groups them by urgency and intent, and sends one digest message. Much quieter than 12 individual Telegram pings overnight. Real-time alerts stay for high-urgency only; the digest handles everything else. Teams that field calls across multiple time zones find this pattern easier to action than real-time noise.

Pattern 3: Contact enrichment before logging

Twilio Webhook → Whisper → GPT-4o-mini → HubSpot Search → [Enrich via Apollo] → HubSpot Note → Sheets

Before creating the note, use an HTTP Request to Apollo or Clearbit to look up the caller's phone number and return company name and job title. Add those fields to the note body and the Sheets row. The note that lands in HubSpot then reads "VP of Operations at Acme Corp, high urgency, callback request" rather than just a raw phone number — which is actually useful context.

n8n nodes you'll use most

Getting started

1. Set up a Twilio number with voicemail enabled. Under Voice Configuration, point the voicemail fallback webhook at your n8n Webhook URL.
2. Create an OpenAI API key with Whisper and Chat Completions scopes enabled.
3. Connect HubSpot via OAuth2 in n8n's credential manager — API key works for free-tier contact + note operations.
4. Create a Google Sheet with columns: Timestamp, From, Duration, Urgency, Intent, Summary, Transcript.
5. Add a Telegram bot token and configure your target chat ID.
6. Build the workflow chain: Webhook → HTTP download → HTTP Whisper → Code parse → HTTP GPT-4o-mini → Code parse → HubSpot → If → Sheets and Telegram in parallel.
7. Test with a real voicemail: call your Twilio number, leave a short message, then watch the n8n execution log. The transcript and HubSpot note should appear within 30 seconds of the recording completing.

For teams applying the same AI triage approach to email inboxes, the patterns in n8n AI email auto-reply map directly — same urgency classification logic, different trigger node. The n8n HubSpot automation post covers the contact-matching and note-creation config in more depth if the CRM side is the trickier part of the build. Start with the VoiceNote template for a pre-wired version, or browse the full n8n template catalog to see what else maps to your current stack.