n8n AI Email Automation Workflow: Full Developer Tutorial (2026)

What You Will Build — Full Pipeline Overview

Here is the complete workflow you will have running by the end of this tutorial. Every box is an n8n node. Data flows left to right.

What this pipeline does, end to end:

1. Monitors your inbox every minute via IMAP
2. Converts HTML email bodies to clean plain text
3. Sends each email to GPT-4o-mini for classification (support, billing, sales, spam, etc.)
4. Routes each email to the correct branch based on its category
5. Generates a context-aware draft reply using a tuned system prompt
6. Either sends the reply automatically (for high-confidence categories) or saves it as a Gmail draft for review

The advanced section (Step 8) adds a Qdrant vector database so the AI can answer questions using your actual company documentation — not just generic knowledge.

Prerequisites & Stack Overview

Cost Reality Check: Running 500 emails/day through GPT-4o-mini for classification + drafting costs approximately $1–3/day in API fees. Self-hosting n8n on a $6/month Hetzner VPS means your total cost is under $100/month for a full AI email automation system — versus $500+/month for SaaS alternatives.

Step 1 — Install n8n Self-Hosted (Docker)

The fastest production-ready setup uses Docker Compose with PostgreSQL for persistent storage. Run these commands on your server or local machine.

1a. Create the project directory

1b. Create the Docker Compose file

1c. Start n8n

Open http://localhost:5678 in your browser. Create your admin account. You are now inside the n8n canvas.

Quick Cloud Alternative: Not ready to self-host? n8n Cloud offers a free trial at n8n.io — skip straight to Step 2. You can always migrate your workflow JSON to self-hosted later.

Step 2 — IMAP Email Trigger Node Setup

This node monitors your inbox and fires the workflow every time a new email arrives. In n8n, click + Add Node → search “Email Trigger (IMAP)”.

Gmail App Password Setup (Important)

Gmail blocks basic IMAP authentication by default. You need an App Password:

• Enable 2-Factor Authentication on your Google account: Go to myaccount.google.com → Security → 2-Step Verification
• Generate an App Password: myaccount.google.com → Security → App Passwords → Select “Mail” → Generate. Copy the 16-character password.
• Use App Password in n8n IMAP credential: Host: imap.gmail.com | Port: 993 | SSL: ON | Password: the 16-character App Password

Step 3 — Email Preprocessing Node

Emails arrive as raw HTML. Sending HTML directly to the AI wastes tokens and confuses the model with tags. This node converts it to clean plain text first.

Add a Code Node (search “Code”) after the IMAP trigger. Set language to JavaScript and paste:

// Clean HTML email body to plain text
// Runs for each email item the IMAP trigger produces

const items = $input.all();

return items.map(item => {
  const html = item.json.text?.html || item.json.text?.plain || "";

    // Strip HTML tags
  let plain = html
    .replace(/<style[^>]*>[\s\S]*?<\/style>/gi, "")
    .replace(/<script[^>]*>[\s\S]*?<\/script>/gi, "")
    .replace(/<[^>]+>/g, " ")
    .replace(/&nbsp;/g, " ")
    .replace(/&amp;/g, "&")
    .replace(/&lt;/g, "<")
    .replace(/&gt;/g, ">")
    .replace(/\s+/g, " ")
    .trim();

    // Truncate to 3000 chars to save tokens
  if (plain.length > 3000) {
    plain = plain.substring(0, 3000) + "... [truncated]";
  }

  return {
    json: {
      ...item.json,
      cleanBody: plain,
      subject: item.json.subject || "(no subject)",
      from: item.json.from?.value?.[0]?.address || "unknown",
      fromName: item.json.from?.value?.[0]?.name || "",
    }
  };
});

Step 4 — AI Email Classifier Node

Add an OpenAI node after the Code node. This node sends each email to GPT-4o-mini and gets back a single category label.

Using Ollama Instead of OpenAI (Free & Private): If you want fully local classification with no API costs, replace the OpenAI node with an Ollama node using llama3.1 or mistral. Accuracy is slightly lower but privacy is complete — nothing leaves your server. Add ollama as a service in your docker-compose.yml.

Step 5 — Switch Router (Branch by Category)

Add a Switch node after the classifier. This routes each email to a different branch based on what the AI returned.

Each output of the Switch node connects to a different branch. The SPAM branch goes directly to an archive action. Other branches go to the AI Draft node with different system prompts.

Step 6 — AI Draft Generator Node

Each non-spam branch gets its own OpenAI node with a system prompt tuned for that category. Here is the configuration for the Support branch — replicate it for other categories, adjusting the system prompt.

System Prompts for Other Branches

Paste these into the matching branch’s OpenAI draft node:

System Prompt — BILLING_QUERY Branch

System Prompt — SALES_INQUIRY Branch

Step 7 — Send Reply or Save as Gmail Draft

After the AI generates the draft, you have two options. Option A saves it as a Gmail draft for human review (recommended to start). Option B sends automatically (use only after testing thoroughly).

Option A — Save as Gmail Draft (Recommended First)

Option B — Send Automatically via SMTP

Do Not Auto-Send Until You Have Tested 50+ Emails: Run the workflow in “Save as Draft” mode for at least one week before enabling auto-send. Read every draft, correct the prompts, and only switch to automatic sending once you are confident the AI handles edge cases correctly.

Step 8 — (Advanced) Add RAG with Qdrant Vector Database

By default, the AI drafts replies using only its training knowledge. With RAG (Retrieval-Augmented Generation), it can also search your company’s actual documentation — FAQs, product specs, pricing tables — before drafting a reply.

What RAG Changes

Without RAG: AI answers from general knowledge → may get product details wrong.
With RAG: AI searches your docs first → answers are specific, accurate, and grounded in your actual information.

8a. Add Qdrant to Docker Compose

8b. The RAG Pipeline in n8n

Between the Switch Router and the Draft Generator, add these three nodes for each branch that needs knowledge-based replies:

• Embeddings OpenAI Node: Takes the cleaned email body and generates a vector embedding using text-embedding-3-small model. This converts the email’s meaning into numbers Qdrant can search.
• Qdrant Vector Store Node — Search: Uses the embedding to search your Qdrant collection for the most relevant documentation chunks. Set Top K to 3 (returns 3 most relevant passages).
• Code Node — Build RAG Context: Combines the retrieved document chunks into a context string that gets prepended to the AI draft prompt — “Based on the following documentation: [chunks] — reply to this email: [email]”

8c. Load Your Documents into Qdrant

Create a separate n8n workflow that runs once (or weekly) to vectorize your documentation:

Step 9 — Error Handling & Slack Alerts

Production workflows need error handling. Without it, a failed OpenAI call silently drops an email and you never know.

9a. Add Error Workflow

In n8n Settings → go to each workflow → set Error Workflow. Create a separate workflow called “Email Automation Error Handler” with:

• Error Trigger Node: This fires whenever any node in your main workflow throws an error.
• Slack / Email Notification Node: Send yourself an alert with the error message, the email subject that failed, and the node name. This gives you visibility without monitoring constantly.
• Fallback: Save Failed Email to Google Sheets: Log the original email to a spreadsheet so you can manually reply to anything the automation could not handle.

9b. Add Retry Logic for OpenAI Calls

OpenAI APIs occasionally time out. In each OpenAI node settings, enable “On Error: Retry” with 2 retries and a 5-second wait.

Step 10 — Production Checklist

Before activating the workflow for real email, go through this checklist:

You Are Live: Once the checklist passes, activate your workflow in n8n (toggle at top right of the canvas). The IMAP trigger will begin polling and your AI email pipeline is running. Monitor the Executions tab for the first 48 hours to catch any edge cases.

Frequently Asked Questions