אחיה כהן

Posted on Mar 20 • Edited on Mar 31 • Originally published at achiya-automation.com

How to Build a Self-Hosted Customer Support Pipeline with n8n, Supabase, and WhatsApp

#automation #n8n #whatsapp #discuss

Why Self-Hosted?

Most customer support automation guides point you to Zendesk, Intercom, or Freshdesk. They're great products — but they come with per-seat pricing that can crush a small business budget. At $49-99/agent/month, a 3-person support team costs $1,800-3,600/year before you even start customizing.

What if you could build the same thing with open-source tools, running on a $20/month VPS?

Here's the architecture I keep coming back to. It handles WhatsApp as the primary channel, stores everything in a real database (not spreadsheets), and automates the entire pipeline from first contact to resolution.

The Stack

Component	Tool	Role	Cost
Automation Engine	n8n (self-hosted)	Workflow orchestration	Free (open-source)
Database	Supabase (self-hosted)	Customer data, ticket tracking	Free (open-source)
Messaging	WhatsApp Business API	Customer communication	~$0.05/conversation
AI	OpenAI API	Intent classification	~$0.01/request
Hosting	Any VPS (Hetzner, DigitalOcean)	Infrastructure	~$20/month

Total: ~$25/month vs $150+/month for SaaS equivalents.

Architecture Overview

Customer (WhatsApp) -> WAHA/Cloud API -> n8n Webhook
    -> Router Node (by intent)
        -> FAQ: AI Response -> Send Reply
        -> Appointment: Check Calendar -> Book -> Confirm
        -> Complaint: Create Ticket -> Notify Team
        -> Sales: Qualify Lead -> CRM Entry -> Assign Agent
    -> Supabase (PostgreSQL)

Step 1: Set Up the WhatsApp Webhook

Configure your WhatsApp API (Meta Cloud API or WAHA for self-hosted) to send incoming messages to an n8n webhook:

// n8n Webhook Node Configuration
{
  "httpMethod": "POST",
  "path": "whatsapp-support",
  "responseMode": "onReceived",
  "responseCode": 200
}

The webhook receives: from (phone number), body (message text), timestamp, and type (text, image, document).

Step 2: Intent Classification with AI

// n8n Function Node
const message = $input.first().json.body;
const prompt = 'Classify this customer message into ONE category: FAQ, APPOINTMENT, COMPLAINT, SALES, OTHER. Message: "' + message + '"';
return [{ json: { prompt, originalMessage: message } }];

This costs ~$0.001/message and runs in under 500ms. Handles Hebrew, English, and mixed languages automatically.

Step 3: Database Schema (Supabase)

CREATE TABLE customers (
  id UUID DEFAULT gen_random_uuid() PRIMARY KEY,
  phone TEXT UNIQUE NOT NULL,
  name TEXT,
  first_contact TIMESTAMPTZ DEFAULT NOW(),
  last_contact TIMESTAMPTZ DEFAULT NOW(),
  total_interactions INT DEFAULT 1
);

CREATE TABLE tickets (
  id SERIAL PRIMARY KEY,
  customer_id UUID REFERENCES customers(id),
  category TEXT NOT NULL,
  message TEXT NOT NULL,
  status TEXT DEFAULT 'open',
  priority TEXT DEFAULT 'normal',
  created_at TIMESTAMPTZ DEFAULT NOW()
);

CREATE TABLE messages (
  id SERIAL PRIMARY KEY,
  customer_id UUID REFERENCES customers(id),
  direction TEXT NOT NULL,
  content TEXT NOT NULL,
  channel TEXT DEFAULT 'whatsapp',
  created_at TIMESTAMPTZ DEFAULT NOW()
);

Step 4: The Router Workflow

Use a Switch node after AI classification:

FAQ: Match FAQ database -> AI response -> Send via WhatsApp
Appointment: Parse date -> Check calendar -> Book -> Confirm
Complaint: Create ticket -> Notify team -> Acknowledge
Sales: Qualify lead -> CRM entry -> Assign agent

Step 5: Upsert Customer Records

// n8n Supabase Node
{
  "operation": "upsert",
  "table": "customers",
  "conflictColumn": "phone",
  "data": {
    "phone": "{{ $json.from }}",
    "name": "{{ $json.profile_name }}",
    "last_contact": "{{ $now }}"
  }
}

Step 6: Automated Follow-ups

Set up n8n Cron triggers for:

24-hour open ticket alerts
Post-resolution satisfaction surveys
48-hour lead re-engagement

Production Considerations

Rate limiting: Start with 250 business-initiated conversations/24h
Error handling: Fall back to templates if AI fails
Monitoring: Use n8n execution logs + alerts for failed webhooks
Backup: Standard pg_dump via n8n cron job

What to Expect

The main benefit of this setup isn't some magical metric — it's that every customer message gets an instant acknowledgment, even at 3am. That alone changes the dynamic. Customers feel heard, and the business owner isn't chained to their phone.

The AI intent classification handles routine questions (hours, pricing, directions) without human involvement. More complex issues get routed to the right person with full context. It's not "80% automation" — it's "the simple stuff happens automatically, and the hard stuff gets escalated properly."

What's Next?

Extend with voice note transcription (Whisper), image processing (vision AI), multi-language support, and analytics dashboards (Metabase).

The key insight: you don't need enterprise software for enterprise-grade support.

Building automation systems for small businesses at Achiya Automation. Try n8n — it's free and open source. We specialize in WhatsApp bots, workflow automation, and AI agents — all self-hosted, no vendor lock-in.

Over to You

If you could only self-host ONE tool in your support stack, which would it be — the automation engine, the database, or the customer inbox?

I chose the inbox first (Chatwoot). That was a mistake. The automation engine should come first because without it, you're just self-hosting a fancy email client.

Here's what I'm still figuring out: at what volume does Chatwoot start struggling? I've seen slowdowns around 10K conversations/month on a 4GB VPS. Anyone running it at higher scale?

Share your self-hosted support stack below — I'll tell you what I'd change. 👇

DEV Community