How to Build an AI Email Auto-Responder with n8n and Ollama (No API Keys Needed)

Published March 24, 2026 · 18 min read · Beginner-friendly

Your inbox is a time sink. The average professional spends over 2 hours per day reading and responding to emails, and most of those replies follow predictable patterns. What if you could build an AI system that reads every incoming email, classifies it by type and urgency, and drafts a context-aware reply — all running on your own machine with zero API costs?

In this n8n Ollama tutorial, you will build exactly that: a self-hosted AI email automation pipeline that connects your inbox to a locally-running AI model. No OpenAI key. No cloud dependency. No per-request charges. Just your hardware, doing the work.

By the end, you will have a working n8n AI workflow that polls your inbox every 5 minutes, classifies each message (support, sales, spam, personal), rates the urgency, and drafts a reply that you can review and send with one click.

Prerequisites

You do not need to be an AI expert. If you have installed software from the terminal before, you can follow this tutorial. Here is what you need:

Time required: approximately 20 minutes for the first setup. After that, the workflow runs unattended.

How the Workflow Works

Before we start building, here is the high-level flow of the local AI email automation system:

1. IMAP Trigger — n8n polls your inbox on a schedule (every 5 minutes by default)
2. AI Classification — Ollama reads the email and returns a structured JSON classification: category, urgency, sentiment
3. Smart Routing — A Switch node routes emails to different handling paths based on the classification
4. AI Reply Drafting — Ollama generates a context-aware reply, tailored to the email category
5. Output — The drafted reply is saved to a Google Sheet, sent via webhook, or placed into your drafts folder

Each step is a node in n8n. The entire pipeline runs locally. Your email content never touches a third-party AI provider.

Step 1: Install Ollama and Pull a Model

# Linux or macOS (one-line install)
curl -fsSL https://ollama.ai/install.sh | sh

# Windows: download the installer from https://ollama.ai/download

ollama pull llama3:8b

curl http://localhost:11434/api/tags

Step 2: Set Up n8n and Verify the Connection

docker run -d --name n8n \
  -p 5678:5678 \
  -v n8n_data:/home/node/.n8n \
  --add-host=host.docker.internal:host-gateway \
  n8nio/n8n

Step 3: Connect Your Inbox with IMAP

The IMAP node will poll for new emails on whatever schedule you set in the workflow settings (we will set it to every 5 minutes later). Each new email arrives as a JSON object with from, subject, text, and date fields.

Step 4: Build the AI Classification Node

This is where the self-hosted AI automation starts. We send the email content to Ollama and ask it to classify the message into structured categories.

{
  "model": "llama3:8b",
  "prompt": "You are an email classification assistant. Analyze this email and return a JSON object.\n\nClassify into exactly ONE category: support, sales, partnership, personal, newsletter, spam\nRate urgency: low, medium, high\nDetect sentiment: positive, neutral, negative\nProvide a 1-sentence summary.\n\nEmail from: {{ $json.from }}\nSubject: {{ $json.subject }}\nBody: {{ $json.text }}\n\nReturn ONLY valid JSON, no other text:\n{\"category\": \"...\", \"urgency\": \"...\", \"sentiment\": \"...\", \"summary\": \"...\"}",
  "stream": false,
  "options": {
    "temperature": 0.1
  }
}

A few things to note about this prompt:

• Temperature 0.1 makes the output nearly deterministic. For classification tasks, you want consistency, not creativity.
• "Return ONLY valid JSON" is important. Without this instruction, the model tends to add explanations before or after the JSON, which breaks parsing.
• The double-brace expressions ({{ $json.from }}) are n8n template expressions that inject the email data into the prompt.

Step 5: Parse and Route by Category

// Extract JSON from the AI response
const aiResponse = $input.first().json.response;
const jsonMatch = aiResponse.match(/\{[\s\S]*\}/);

if (!jsonMatch) {
  throw new Error('AI did not return valid JSON. Raw response: ' + aiResponse);
}

const classification = JSON.parse(jsonMatch[0]);

return [{
  json: {
    ...classification,
    original_from: $('IMAP Email').first().json.from,
    original_subject: $('IMAP Email').first().json.subject,
    original_body: $('IMAP Email').first().json.text,
    processed_at: new Date().toISOString()
  }
}];

Step 6: Draft Smart Replies with AI

Now we build the reply-drafting nodes. Each category gets its own HTTP Request node calling Ollama, but with a different prompt tailored to that type of email.

{
  "model": "llama3:8b",
  "prompt": "Draft a professional support reply to this email.\n\nFrom: {{ $json.original_from }}\nSubject: {{ $json.original_subject }}\nBody: {{ $json.original_body }}\n\nGuidelines:\n- Be empathetic and acknowledge the issue\n- Provide specific, actionable troubleshooting steps\n- Offer to escalate if the steps don't resolve it\n- Keep it under 150 words\n- Professional but warm tone\n\nWrite ONLY the reply body, no subject line or signature.",
  "stream": false,
  "options": { "temperature": 0.5 }
}

{
  "model": "llama3:8b",
  "prompt": "Draft a reply to this sales inquiry.\n\nFrom: {{ $json.original_from }}\nSubject: {{ $json.original_subject }}\nBody: {{ $json.original_body }}\n\nGuidelines:\n- Thank them for their interest\n- Answer their specific question if one was asked\n- Briefly highlight key benefits relevant to their inquiry\n- Include a clear call to action (book a demo, start a trial, etc.)\n- Confident but not pushy\n- Keep it under 150 words\n\nWrite ONLY the reply body.",
  "stream": false,
  "options": { "temperature": 0.5 }
}

For the generic/default reply, use a neutral prompt that acknowledges the email and says you will follow up shortly. The higher temperature (0.5) gives the replies more natural variation compared to the classification step.

Step 7: Send or Queue the Drafted Reply

Going to Production

Once you have verified the workflow handles your emails correctly, here are some improvements for production use:

Add Error Handling

Wrap the AI classification and reply nodes in n8n's error workflow feature. If Ollama times out or returns garbage, the error handler can log the failure and forward the email to you unprocessed rather than losing it.

Filter Before Classification

Add a filter node before the AI step to skip emails you never want to process: automated notifications from services, calendar invites, delivery receipts. This saves processing time and keeps your logs clean.

Log Everything

Add a Google Sheets or database node at the end of every path to log the email address, classification result, and draft reply. This gives you an audit trail and lets you measure classification accuracy over time.

Model Selection for Production

For email classification specifically, llama3:8b hits the sweet spot of speed and accuracy. If you need faster processing on CPU-only hardware, phi3:mini (3.8B parameters) handles classification well at roughly 3x the speed. For reply drafting where quality matters more, stick with the 8B model or go larger if your hardware allows.

Troubleshooting

Connection Issues

• "Connection refused" from n8n to Ollama — If n8n runs in Docker, you must use http://host.docker.internal:11434 instead of localhost. If n8n is installed via npm on the same machine, localhost works fine.
• Ollama not running — Run ollama serve in a terminal. On Linux, you may want to set it up as a systemd service so it starts on boot.

AI Output Issues

• AI returns text instead of JSON — Strengthen the prompt: add "You MUST return ONLY a JSON object. No explanations, no markdown, no code fences." Also ensure temperature is below 0.3 for classification.
• Classification is inconsistent — Lower the temperature to 0.05. Add 2-3 example classifications to your prompt (few-shot prompting). This dramatically improves consistency.
• Replies are too generic — Add specific context to the reply prompt: your company name, product details, common FAQ answers. The more context the model has, the better the replies.

Performance Issues

• Timeout errors — Increase the HTTP Request timeout to 180000ms (3 minutes). First-run requests are slower because the model is loading into memory.
• Slow on CPU — Use a smaller model like mistral:7b or phi3:mini. Even a modest GPU (RTX 3060, 12 GB VRAM) makes a 5-10x speed difference.
• High memory usage — Ollama loads the model into RAM. Close other memory-intensive applications. If you have exactly 8 GB RAM, use a smaller model and close your browser while running the workflow.

IMAP Issues

• Gmail "Application-specific password required" — Enable 2FA on your Google account, then generate an App Password at myaccount.google.com/apppasswords.
• No emails being fetched — Check that the mailbox name is correct (INBOX is case-sensitive for some providers). Test your credentials with a desktop email client first.

Next Steps

You now have a working local AI email automation pipeline. Here are a few ways to extend it:

• Add sentiment-based priority: Route high-urgency, negative-sentiment emails to a Slack notification so you respond to angry customers immediately.
• Build a feedback loop: Track which drafts you edit heavily before sending. Use those corrections to refine your prompts over time.
• Multi-language support: Add "Detect the language of the email and reply in the same language" to your reply prompt. Models like llama3 handle this well.
• Connect to a CRM: Use the classification data to auto-create or update contacts in your CRM based on email category and sentiment.

Related Tutorials

• How to Build Self-Hosted AI Workflows with n8n and Ollama (2026 Guide)
• n8n + Ollama Tutorial: Automate Everything with Local AI
• How to Automate Email Responses with n8n and Ollama

Free Samples

Try these open-source workflows to see the quality before buying the pack:

• Free: AI Social Media Content Generator
• Free: AI Blog Writer Pipeline
• Free: AI Email Auto-Responder (the workflow from this tutorial)