Use VoxFlow as the AI brain, n8n as the business logic.

The separation of concerns

VoxFlow does one thing: it makes a phone call into an AI conversation. It doesn’t know about your CRM, your calendar, your email templates, or your business rules. That’s deliberate.

flowchart LR
    Caller([Caller]) --> VoxFlow[VoxFlow
AI]
    VoxFlow --> N8N[n8n workflow]
    N8N --> CRM[CRM lookup]
    N8N --> Cal[Calendar booking]
    N8N --> Email[Email confirmations]
    N8N --> Slack[Slack notifications]

VoxFlow is the mouth and ears. n8n is the hands.

The contract: three webhook routes

VoxFlow posts to a single N8N_WEBHOOK_URL, distinguishing intent with a route field:

route When Payload Expected reply
"1" New call starts {number, data: "empty"} {firstMessage: "<greeting>"}
"2" Call ended {number, data: "<full transcript>"} (ignored)
"3" AI booked a meeting {number, data: "<JSON details>"} {message: "<confirmation>"}

This is the entire integration surface. Implement these three routes and you’ve integrated VoxFlow with whatever system you want.

The minimal n8n workflow

flowchart TD
    WH[Webhook Trigger
POST /webhook/abc123] --> SW{Switch on
$json.route}
    SW -- "1" --> R1[HTTP Request: CRM lookup by phone]
    R1 --> R1b["Set: {firstMessage: 'Hi , ...'}"]
    SW -- "2" --> R2[Postgres: INSERT INTO call_log]
    R2 --> R2b[Optional: Slack notification]
    SW -- "3" --> R3[Google Calendar: Create event]
    R3 --> R3b[Email: Send confirmation]
    R3b --> R3c["Set: {message: 'Booked for '}"]

That’s it. Three branches. No code.

Why personalize the greeting?

Route "1" fires at the start of every call. If you have any record of this phone number, you can greet them by name:

1
2
Anonymous caller:   "Hello, this is Sara. How can I help?"
Returning customer: "Hi John, welcome back! Calling about your appointment Tuesday?"

The second one cuts 15 seconds off the call by skipping identity verification. Over 1000 calls/month that’s hours of phone time saved.

If n8n returns nothing or fails, VoxFlow falls back to DEFAULT_FIRST_MESSAGE. The call never breaks.

Why ship the transcript?

Route "2" is fire-and-forget. The call has ended; we’re just archiving. Common downstream uses:

  • Compliance log — financial / healthcare verticals require recordings or transcripts
  • Quality assurance — sample 5% of calls into a Slack channel for human review
  • Lead scoring — feed transcript into another LLM for sentiment / intent classification
  • CRM enrichment — append to the contact’s activity timeline

Why route meeting bookings through n8n?

You could have VoxFlow call Google Calendar directly. Don’t.

Reasons to route through n8n instead:

  • Calendar choice — different locations, doctors, salespeople each have different calendars
  • Conflict checks — n8n can query for double-bookings and respond {message: "That slot is taken, how about 3pm?"} which the AI will read aloud
  • Multi-system writes — same booking might create a Salesforce event + send an SMS + update internal ops doc
  • Easier to change — your scheduling logic changes more often than your AI agent

The AI doesn’t care. From its perspective, schedule_meeting() returned a message. It reads that message to the caller.

A complete n8n flow for a dental clinic

flowchart TD
    WH[Webhook trigger] --> SW{Switch on route}

    SW -- "1: greeting" --> G1[HubSpot: find contact by phone]
    G1 --> G2{Found?}
    G2 -- yes --> G3[Build personalized greeting]
    G2 -- no --> G4[Return generic greeting]

    SW -- "2: transcript" --> T1[HubSpot: append to timeline]
    T1 --> T2[GPT-4: extract intent + sentiment]
    T2 --> T3{Sentiment
negative?}
    T3 -- yes --> T4[Slack alert to manager]

    SW -- "3: booking" --> B1[Google Calendar: check availability]
    B1 --> B2{Available?}
    B2 -- yes --> B3[Create event + send SMS]
    B2 -- no --> B4[Return alt slot to AI]

Zero lines of Python written. The receptionist now talks to HubSpot and books appointments on a real calendar.

Authentication

n8n webhooks support HMAC signatures. VoxFlow doesn’t sign requests today, so for production you should either:

  • Put n8n behind a reverse proxy with IP allowlisting (VoxFlow’s outbound IP)
  • Add Authorization headers in send_to_webhook() and validate in n8n
  • Use n8n’s built-in webhook auth (username/password or header token)

Takeaway

VoxFlow + n8n is two specialists doing what they’re each good at. Don’t put business logic in your AI middleware; put it where business analysts can edit it themselves.