How to Automate Customer Support Email Triage with n8n

If you run an online store, your support inbox is a mixed pile: order questions, address changes, refund demands and the occasional angry all-caps message — all looking equally urgent. Triage is the unglamorous work of reading every email just to decide what it is. That part is perfect for automation. Here's how to do it with n8n and an AI model, in a way that never sends anything on your behalf.

What "triage automation" actually does

For every new email that arrives, the workflow:

1. Classifies it into a fixed set of categories — for a store, a good set is: Order Status, Shipping & Delivery, Refund, Return & Exchange, Product Question, Payment & Billing, Complaint, Other.
2. Tags a priority (High / Normal / Low), so a chargeback threat never sits behind a "thanks!" email.
3. Writes a one-line summary, so you can scan fifty emails in a minute.
4. Logs it to a spreadsheet — date, customer, subject, category, priority, summary. Your inbox, pre-sorted.

Notice what's not on the list: replying. Triage is read-only automation. It's the safest place to start, and for many small stores it's already 80% of the time sink.

The workflow, node by node

In n8n, the whole thing is five building blocks:

• Gmail Trigger — fires whenever a new email lands in your support inbox.
• Set node — your store settings (just your store name in the lite version).
• Code node — builds the AI prompt: the customer's email plus a strict instruction to answer with a JSON object containing category, priority and summary.
• HTTP Request node — sends the prompt to an AI API. OpenAI's gpt-4o-mini is a good default (a fraction of a cent per email); any OpenAI-compatible API works, including Zhipu GLM.
• Code + Google Sheets nodes — parse the JSON (with a safe fallback if the model ever returns something unparseable) and append one row to your log sheet.

Two details make this reliable in practice. First, force JSON mode on the AI call and validate the category against your fixed list — free-text answers will drift. Second, make the fallback boring: if parsing fails, log the email as Other / High so a human looks at it, instead of letting the workflow error out silently.

What it costs

n8n is free if you self-host it. The AI call costs well under $0.01 per email with a mini-class model. For a store getting 30 support emails a day, that's roughly a coffee per month — for an inbox that sorts itself.

When you're ready: drafts, not sends

The natural next step is letting the AI also draft replies — saved to your Gmail Drafts, written in your store's voice and in the customer's language — while a human still reviews and presses Send. Add Telegram alerts for the emails flagged as urgent, and you have a support inbox that effectively runs its own first pass. The rule that keeps it safe: the AI may prepare anything, but it never invents order numbers, tracking data or refund approvals, and it never sends.

More guides: 5 customer support emails you can safely automate (and 3 you shouldn't).