n8n WhatsApp AI Agent: Build a 24/7 Auto-Responder (Step-by-Step)

Quick answer: Connect the WhatsApp Trigger node in n8n to your Meta developer app, route each inbound message to an AI model with a constrained system prompt, and send the reply back with the WhatsApp node. Add an escalation branch so unrecognized questions go to a human. The result answers routine questions in seconds, at any hour, without you touching your phone.

The build takes under an hour once your WhatsApp Business Cloud API access is set up. Here is the whole path.

What You Need Before Starting

The Meta setup is the fiddly part: create an app of type Business, add the WhatsApp product, and note the phone number ID and permanent access token. n8n's credential dialog asks for exactly these.

Step 1: Receive Messages With the WhatsApp Trigger

Add the WhatsApp Trigger node and select the messages event. n8n gives you a webhook URL; paste it into your Meta app's WhatsApp webhook configuration together with the verify token n8n shows.

Send a test message to your business number and confirm the execution fires. The customer's text arrives in the node output along with their WhatsApp ID, which you will use for the reply.

Step 2: Generate the Answer With a Constrained AI Step

Add your AI model node and build a system prompt with three parts: who the assistant is, what it knows, and when it must hand off. A working skeleton:

You are the assistant for [business name]. Answer customer questions
using ONLY the information below. Keep replies under 100 words,
friendly and specific.

BUSINESS INFORMATION
Opening hours: ...
Services and prices: ...
Location and parking: ...
Booking link: ...

RULES
If the question is not covered above, or the customer is upset,
or it involves payments or complaints, reply exactly with: HANDOFF

Pass the inbound message text as the user message. The HANDOFF sentinel is the load-bearing trick: it turns fuzzy AI judgment into a value your workflow can route on. This pattern of keeping the model inside a durable, observable flow is the discipline we describe in AI steps inside n8n, Make, and Zapier.

Step 3: Route the Response

Add an IF node checking whether the model output contains HANDOFF.

Normal branch: a WhatsApp node with the Send Message operation returns the model's answer to the customer's WhatsApp ID. Response time from trigger to reply is a few seconds.

Escalation branch: notify a human with the customer's number and full question, then send the customer a short holding message saying a teammate will reply shortly. A Slack alert works perfectly here, and the wiring is the same as in our Sheets to Slack tutorial.

Step 4: Log Everything

Add a final step that appends every conversation to a Google Sheet or database: timestamp, customer number, question, answer, and whether it escalated. The log is how you audit the bot's answers, spot the questions it cannot handle, and improve the prompt weekly. An unlogged AI agent talking to customers is a liability, not an asset.

The Starter Template

Import and rewire with your own credentials:

{
  "name": "WhatsApp AI Auto-Responder",
  "nodes": [
    {
      "parameters": { "updates": ["messages"] },
      "id": "wa-trigger-1",
      "name": "WhatsApp Trigger",
      "type": "n8n-nodes-base.whatsAppTrigger",
      "typeVersion": 1,
      "position": [240, 300]
    },
    {
      "parameters": {
        "modelId": { "__rl": true, "mode": "list", "value": "gpt-4o-mini" },
        "messages": {
          "values": [
            { "role": "system", "content": "REPLACE_WITH_SYSTEM_PROMPT" },
            { "role": "user", "content": "={{ $json.messages[0].text.body }}" }
          ]
        }
      },
      "id": "ai-1",
      "name": "AI Answer",
      "type": "@n8n/n8n-nodes-langchain.openAi",
      "typeVersion": 1.8,
      "position": [480, 300]
    },
    {
      "parameters": {
        "conditions": {
          "string": [{ "value1": "={{ $json.message.content }}", "operation": "notContains", "value2": "HANDOFF" }]
        }
      },
      "id": "if-1",
      "name": "Needs Human?",
      "type": "n8n-nodes-base.if",
      "typeVersion": 2,
      "position": [720, 300]
    },
    {
      "parameters": {
        "operation": "send",
        "phoneNumberId": "REPLACE_WITH_PHONE_NUMBER_ID",
        "recipientPhoneNumber": "={{ $('WhatsApp Trigger').item.json.messages[0].from }}",
        "textBody": "={{ $json.message.content }}"
      },
      "id": "wa-send-1",
      "name": "Reply on WhatsApp",
      "type": "n8n-nodes-base.whatsApp",
      "typeVersion": 1,
      "position": [960, 220]
    }
  ],
  "connections": {
    "WhatsApp Trigger": { "main": [[{ "node": "AI Answer", "type": "main", "index": 0 }]] },
    "AI Answer": { "main": [[{ "node": "Needs Human?", "type": "main", "index": 0 }]] },
    "Needs Human?": { "main": [[{ "node": "Reply on WhatsApp", "type": "main", "index": 0 }]] }
  },
  "active": false,
  "settings": { "executionOrder": "v1" }
}

Wire your escalation branch from the second output of the IF node to your Slack or email node, and add your logging step at the end of both branches.

Two Cautions Before You Go Live

Privacy. Customer messages are personal data. Host thoughtfully, minimize what your execution logs retain, and know your obligations; our n8n GDPR guide covers exactly this decision.

Scope. Launch with a narrow, honest bot that answers ten questions well and hands off everything else. Widen its knowledge weekly based on the log. Bots fail publicly when they are allowed to improvise on day one.