Automate Email Triage with n8n and OpenAI: A Step-by-Step Tutorial

You are an email classifier. Classify the following email into exactly one of these categories: WORK, PERSONAL, SPAM, NEWSLETTER, SUPPORT, BILLING. Respond with only the category name, nothing else.

{{ $json["text"] }}

{{ $json["choices"][0]["message"]["content"].trim() }}

Step 4: Route Emails by Category Using a Switch Node

Now that you have the category, you need to send the email to the correct branch of the workflow. The Switch node in n8n allows you to route data based on a condition.

1. Add a new node after the Set node. Search for "Switch" and select the "Switch" node.
2. Configure the node:
   • Mode: use "Expression"
   • Expression: {{ $json["category"] }}
   • Output Type: "String"
   • Fallback Output: create a default output for uncategorized emails (e.g., label them as "UNCATEGORIZED")
3. Click "Add Routing" for each category you want to handle. The tutorial uses six categories: WORK, PERSONAL, SPAM, NEWSLETTER, SUPPORT, BILLING.
4. For each routing, set the Value to the exact category name (e.g., "WORK"). The values must match the output from the LLM exactly, including case. If the LLM returns "work" (lowercase), the routing will fail. To be safe, you can use the expression {{ $json["category"].toLowerCase() }} in the Switch node and ensure your routing values are also lowercase.

Why this matters: The Switch node is the decision point. It examines the category and sends the email down the correct path. Without it, you would have to manually check each email. The fallback output catches any emails that the LLM could not classify, preventing them from being lost.

Expected output: When you test the Switch node with a sample email that has a category of "WORK", the node should route the email to the output labeled "WORK". You can verify this by looking at the node's output panel, which shows which branch was taken.

Step 5: Apply Gmail Labels Automatically

For each category branch, you need a Gmail node that applies the corresponding label. You will repeat this step for each category.

1. Add a new node after the Switch node's "WORK" output. Search for "Gmail" and select the "Gmail" node.
2. Choose the operation "Label".
3. Click "Create New Credential" for Gmail. You will need to authorize n8n to access your Gmail account. Follow the OAuth flow in the browser.
4. Configure the node:
   • Operation: "Add Label"
   • Message ID: use the expression {{ $json["messageId"] }} (this field comes from the Email Trigger node; if your trigger node uses a different field name, adjust accordingly)
   • Label Name: enter the exact name of the Gmail label you created earlier, e.g., "WORK"
5. Repeat this for each category branch (PERSONAL, SPAM, NEWSLETTER, SUPPORT, BILLING). You can duplicate the Gmail node and change the label name for each one.

Why this matters: This is the final action that makes your inbox organized. Instead of manually sorting emails, the workflow automatically applies labels as soon as an email arrives. You can then create Gmail filters or views based on these labels.

Expected output: When you test the workflow with a real email, you should see the label appear in Gmail within a few seconds. Open Gmail, find the email, and check that the label is applied.

Alternative approach: If you prefer not to use Gmail labels, you can use other actions like forwarding the email to a different address, sending a Slack notification, or adding the email to a Google Sheet. The tutorial focuses on Gmail labels because they are the most direct way to organize your inbox.

Step 6: Handle the Fallback (Uncategorized Emails)

Not every email will be cleanly classified. The LLM might return an unexpected category, or the email might be in a language the model does not handle well. The fallback output from the Switch node catches these cases.

1. Add a node to the fallback output of the Switch node. You can use a "Gmail" node with the label "UNCATEGORIZED" or simply send a notification to yourself.
2. Alternatively, use an "Email" node to send yourself a summary of the uncategorized email so you can manually classify it.
3. Configure the Email node:
   • From: your email address
   • To: your email address
   • Subject: Uncategorized email: {{ $json["subject"] }}
   • Body: The following email could not be classified: {{ $json["text"] }}

Why this matters: The fallback ensures that no email is silently dropped. You will know immediately if the LLM fails to classify an email, and you can take manual action or adjust the prompt.

Step 7: Add Error Handling (Optional but Recommended)

n8n workflows can fail for many reasons: API rate limits, network timeouts, malformed email data. Adding error handling makes the workflow more robust.

1. Click on the OpenAI node and go to the "Error Handling" tab.
2. Enable "Continue on Fail".
3. Add a new node connected to the error output of the OpenAI node. Use a "Set" node to set a default category of "UNCATEGORIZED" when the LLM call fails.
4. Connect this error path to the same Gmail labeling logic you built for the fallback.

Why this matters: Without error handling, a single failed API call can stop the entire workflow. With "Continue on Fail" enabled, the workflow proceeds even if the LLM is temporarily unavailable, and the email still gets a label (even if it is just "UNCATEGORIZED").

Verifying It Works

To verify the entire workflow is functioning correctly, send a test email to your Gmail account from a different address. The email should be processed within the polling interval you set (e.g., 5 minutes).

1. Send an email with a clear subject and body that matches one of your categories. For example, send an email with the subject "Invoice for March Services" and body "Please find attached the invoice for March." This should be classified as BILLING.
2. Wait for the polling interval to pass, or manually trigger the workflow by clicking "Execute Workflow" in the n8n editor.
3. Open Gmail and check that the email has the BILLING label applied.
4. Repeat with emails that should be classified as WORK, PERSONAL, SPAM, NEWSLETTER, and SUPPORT.
5. Check the workflow execution history in n8n. Each execution should show a green checkmark. If any execution shows an error, click on it to see which node failed and why.

Expected result: All test emails should have the correct Gmail labels applied within minutes. The workflow should run silently in the background without any manual intervention.

Common Problems

Problem: The LLM returns a category that does not match any routing value.

Cause: The LLM might return a slightly different string, such as "Work" instead of "WORK", or it might return a category you did not define.

Fix: In the Set node where you extract the category, add .toUpperCase() to the expression: {{ $json["choices"][0]["message"]["content"].trim().toUpperCase() }}. This ensures the category is always uppercase. Also, ensure your Switch node routing values are all uppercase.

Problem: The Email Trigger node fails with "Connection refused" or "Authentication failed".

Cause: Incorrect credentials, IMAP not enabled, or an App Password not generated.

Fix: Verify that IMAP is enabled in Gmail settings. Generate a new App Password if using 2FA. Ensure the password is entered correctly without extra spaces.

Problem: The OpenAI node returns a 429 error (rate limit exceeded).

Cause: You have exceeded your API rate limit or token quota.

Fix: Add a "Wait" node between the Email Trigger and OpenAI nodes. Set the wait time to a random value between 1 and 5 seconds. This spaces out API calls and reduces the chance of hitting rate limits. Alternatively, upgrade your OpenAI plan for higher limits.

Problem: The Gmail node cannot find the message ID.

Cause: The Email Trigger node may use a different field name for the message ID. In some versions of n8n, the field is called id instead of messageId.

Fix: Check the output of the Email Trigger node to see the exact field name. Update the expression in the Gmail node accordingly. Common field names include id, messageId, uid, or gmailMessageId.

Problem: The workflow runs but no labels are applied.

Cause: The Gmail labels may not exist, or the Gmail node is not properly authorized.

Fix: Create the labels in Gmail manually before running the workflow. Go to Gmail Settings > Labels > Create new label. Ensure the label names match exactly what you used in the Gmail node. Re-authorize the Gmail credential in n8n if necessary.

Problem: The LLM classifies everything as SPAM.

Cause: The prompt is too vague or the model is not given enough context.

Fix: Improve the system message. Add examples of what constitutes each category. For instance:

You are an email classifier. Classify the following email into exactly one of these categories: WORK, PERSONAL, SPAM, NEWSLETTER, SUPPORT, BILLING.

Examples:
- "Meeting at 3pm to discuss Q4 results" -> WORK
- "Your Amazon order has shipped" -> PERSONAL
- "You won a free iPhone!" -> SPAM
- "Weekly tech newsletter" -> NEWSLETTER
- "I need help resetting my password" -> SUPPORT
- "Invoice for February services" -> BILLING

Respond with only the category name.

Problem: The workflow processes the same email multiple times.

Cause: The Email Trigger node is set to poll frequently, and the email is not marked as "seen" after processing.

Fix: In the Email Trigger node, enable the option "Mark as Seen" or "Delete After Processing". This ensures that once an email is processed, it is not picked up again on the next poll.

Next Steps

1. Add more categories: Extend the prompt to include additional categories like SOCIAL, URGENT, or FOLLOW_UP. Update the Switch node and add corresponding Gmail labels.
2. Integrate with a CRM: Instead of just labeling emails, use the n8n HTTP Request node to create a contact or log an activity in your CRM (e.g., HubSpot, Salesforce) when an email is classified as SUPPORT or BILLING.
3. Send notifications: Add a Slack or Telegram node to send you a notification when an important email (e.g., BILLING or SUPPORT) arrives. This ensures you never miss a critical message.
4. Summarize newsletters: For emails classified as NEWSLETTER, use another OpenAI node to generate a one-sentence summary and store it in a Google Sheet for later review. This turns your newsletter clutter into a searchable archive.
5. Implement zero-retention AI: If privacy is a concern, switch to an OpenAI-compatible provider like Regolo AI that offers a zero-retention policy. According to the tutorial, Regolo AI does not store any data sent through its API, making it GDPR-compliant by default. You only need to change the Base URL and API key in the OpenAI node credentials.