Auto-tag and route support tickets by topic

01/The build

How to build it

1. 01

   Define the topic taxonomy against your real inbox — on paper first

   Needs a human

   Routing is only as good as the buckets you route into, so write them before you build. Open the last few hundred tickets and group them into the handful of topics that actually recur — for most DTC stores that's: order status / WISMO, returns & exchanges, order change or cancel, damaged / wrong / missing item, product or pre-sale question, billing & payment, wholesale / B2B, and subscription (if you run Recharge). Keep it under ten. For each topic, write down the destination it routes to (a team, an assignee group, or a Gorgias View) and the tag string. Resist a 25-topic taxonomy — every extra bucket is a new way to mis-route, and topics nobody routes on are just noise.

2. 02

   Fire on every new inbound ticket from Gorgias

   Runs itself

   In Gorgias, create a Rule (or an HTTP Integration) with the trigger 'Ticket created' that POSTs the ticket payload — ticket ID, channel (email / chat / Instagram / etc.), requester email, subject, first message body, and whether an order is already linked — to a webhook in n8n. Fire on inbound customer messages only, not on agent replies, so you're never classifying your own team's words. Creation-time is enough for routing; you don't need to re-run on every follow-up the way an escalation rule would.

3. 03

   Run the cheap deterministic pass first

   Runs itself

   Most tickets don't need an LLM. In n8n, classify on free, reliable signals before spending a token: subject and body keyword matches ('return', 'refund', 'where is', 'tracking', 'wholesale', 'broken', 'damaged', 'cancel', 'invoice'), the channel it came in on, and whether Gorgias has already linked a Shopify order. Keep the keyword-to-topic map in one place so it's editable without touching the flow. A clear hit here assigns the topic outright and skips the model entirely — this is what keeps the build fast and cheap at volume.

4. 04

   Fall back to a constrained LLM classification for the rest

   Runs itself

   For tickets the keyword pass can't confidently place ('it's been forever', 'this isn't what I expected', 'do you ship to Canada?'), send the message body plus your fixed taxonomy to an LLM node and ask for a single label and a confidence score. Constrain the output to your exact topic enum so the model can't invent a category — it must return one of your buckets or the literal 'other'. Pass only the message text and the list of allowed topics; you don't need customer PII in the prompt to classify a topic.

5. 05

   Gate on confidence — 'needs-triage' is the safety net

   Runs itself

   This step is what keeps the whole thing trustworthy. If the LLM's confidence is below your threshold, or it returns 'other', or two topics tie, do not route — tag the ticket 'needs-triage' and leave it in the general queue for a human. A confidently mis-routed ticket is worse than an untriaged one, because the customer waits on the wrong person before anyone notices. Set the threshold deliberately and tune it from the weekly review; start cautious and loosen as you trust it.

6. 06

   Detect language and fold it into the route

   Runs itself

   If you sell internationally, a perfectly-tagged returns ticket in Spanish still routes wrong if it lands with an English-only agent. Run a lightweight language-detection step on the message body and treat language as a second routing dimension alongside topic. A Spanish returns ticket routes to the Spanish-speaking team's returns view; if you have no native speaker for a language, route it to a dedicated 'translation-needed' queue rather than dropping it on someone who can't read it.

7. 07

   Write the tag back to Gorgias — idempotently

   Runs itself

   Apply the topic tag (and the language tag) via the Gorgias API. Before writing, read the ticket's existing tags and skip if your topic tag is already present, so a reprocessed or re-opened ticket never gets double-tagged or bounced between routes. Use one canonical topic tag per ticket for routing — add secondary descriptive tags if you like, but route on exactly one — and namespace them (e.g. 'topic:returns', 'lang:es') so they're easy to filter and report on later without colliding with tags your team adds by hand.

8. 08

   Route to the team or view, then review the misses weekly

   Needs a human

   Map each topic tag to its destination from step 1: set the ticket's assignee team, or move it into the matching Gorgias View, so the right specialist sees it first instead of fishing it out of the general pile. Route by team or view rather than a single named agent by default, so a route never dead-ends on someone who's out. Then close the loop: log every ticket's predicted topic and watch which ones a human reassigns — those reassignments are your ground truth. Spend ten minutes a week moving the misses into the keyword pass or tightening the taxonomy. A router nobody tunes drifts into noise; the weekly pass is what keeps agents trusting the routes instead of re-checking them.

02/The stack

What it’s built on

Gorgias

The helpdesk where tickets land and get worked. Its Rules engine fires the webhook on ticket creation, its tags carry the topic and language labels, and its Views and team assignment are the destinations you route into. It also already links each ticket's Shopify order, which the deterministic pass uses as a free signal.Gorgias can do simple keyword-based tagging and assignment in its own Rules on any plan, and offers AI intent detection on higher / add-on tiers. This recipe is worth it when you want a custom taxonomy with an LLM fallback and a confidence gate that you own — tags and logic that live in your n8n flow and would survive a move to another helpdesk, not locked inside one vendor's AI add-on.

n8n

The classifier and router. It receives the Gorgias webhook, runs the keyword pass, calls the LLM only for the leftovers, applies the confidence gate and language detection, then writes the tags back and sets the route via the Gorgias API.n8n is open-source and self-hostable for free, so the classification logic — and the LLM key it calls — stay under your control; their cloud tier just saves you running the server yourself. Zapier works for the keyword-only version but gets awkward once the LLM branch and confidence gate come in.

03/Questions

Before you build