Triage ecommerce support tickets with AI: label, look up, draft and route
Updated · By Ecomsellertool editorial team
To triage ecommerce support tickets with AI, let a model label each email or chat by intent, such as where is my order, return, damaged item, address change or wholesale, and score its confidence. Rules then fetch the order and tracking, a draft reply is written from your templates, and the ticket goes to the right person. The model never refunds, cancels or promises anything.
- Gorgias detects an intent and a sentiment on incoming messages on all Helpdesk plans, and leaves the field blank when it cannot assign one confidently (checked October 3, 2026).
- Zendesk's intelligent triage gives each prediction a confidence field on Suite and Support Professional plans and above; using it in workflows requires the Copilot add-on.
- Shopify Inbox, a free app from Shopify, applies topic labels such as Order status, Edit orders, Returns/Delivered orders and Wholesale only when confidence is high, and only in English.
- Here the model labels and drafts, rules fetch every order fact, and only an in-transit where-is-my-order reply sends without a person.
The step map: who runs each step
Trigger: A customer message arrives: an email to your support address, a chat, or a contact-form submission that becomes a ticket in your helpdesk.
| Step | What happens | Who runs it |
|---|---|---|
| 1 | The new message becomes a ticket; the workflow strips signatures and quoted replies, then finds the customer and their recent orders by email address, phone or an order number in the text. | Rule |
| 2 | An AI model labels the ticket with one intent (where is my order, return, damaged item, address change, wholesale or other), a confidence score, any order number it mentions and the customer's tone. | AI model |
| 3 | Rules fetch the facts that intent needs from Shopify and the carrier: fulfillment status, tracking number and link, estimated delivery date, delivery date and whether the return window is open. | Rule |
| 4 | A routing rule assigns the ticket to a queue and a person by intent, tone, order value and language, and sets its priority. | Rule |
| 5 | The model drafts a reply from your approved templates and policies, using only the facts the rules fetched. It never offers a refund, discount, replacement or a date the carrier did not give. | AI model |
| 6 | A send gate checks the intent, the confidence score, a single matched order and the auto-send list. Only a where-is-my-order reply with in-transit tracking may send on its own; everything else waits as a draft. | Rule |
| 7 | A person reads the draft, edits it and decides anything that costs money or changes an order: refunds, replacements, reshipments, cancellations, address changes and wholesale prices. | Person |
| 8 | Every label, fetched fact, draft and edit is logged, and a weekly sample of sent and edited tickets is checked against the model's labels to tune the threshold. | Rule |
5 rule steps · 2 AI model steps · 1 person step. A rule follows fixed logic you approve; an AI model handles messy input and its output is checked by rules; a person makes the calls listed below.
What stays with a person
- Every refund, replacement, reshipment, discount or cancellation, whatever the ticket asks for
- Address changes, checked against the fulfillment status before the warehouse ships
- Damaged or wrong items, with the customer's photos and the order attached to the draft
- Tickets below the confidence threshold, with no single matched order, or with an angry tone
- Wholesale and bulk enquiries, routed to the person who sets trade prices
Which tools can run it
- Your helpdesk's own triage: Gorgias intents and rules, Zendesk intelligent triage or Shopify Inbox topic labels
- Zapier, Make or n8n with an AI model step, for one inbox, one store and a few templates
- Custom code on your accounts, for several stores or marketplaces, order data outside Shopify or strict auto-send rules
Systems it connects
- Helpdesk: Gorgias, Zendesk or Shopify Inbox
- Support email inbox
- Shopify orders and fulfillments
- Carrier tracking
- 3PL or warehouse system
- Reply templates and store policies
- Slack or email for escalations
Build size: Medium
The model step is small; the work is matching tickets to orders, the reply templates and a send gate your team trusts.
Most of a store's support inbox asks the same few things: where is my order, how do I return this, it arrived broken, can you change my address. Someone opens each ticket, looks up the order in Shopify, copies the tracking link and types a reply. An AI model can take over the reading and the typing. It should not take over the decisions. This recipe splits the work so the model labels and drafts, rules fetch the facts, and a person keeps everything that costs money. It is one page of our automation library.
Which tickets can AI handle, and which need a person?
Start with this table, because every rule in the workflow comes from it. Each row is a ticket type. The rule column is what the workflow looks up without any AI. The model column is the only work the model does. The person column is what never leaves a human. The last column is the one to argue about with your team before anything is built.
| Ticket type | What the rule fetches | What the model does | What the person decides | Auto-send allowed? |
|---|---|---|---|---|
| Where is my order | The matched order, its fulfillment status, tracking number and link, estimated delivery date | Labels the intent; drafts a reply that states the status and the carrier's estimate | Late, lost or stuck parcels: whether to reship, refund or open a carrier claim | Yes, only when one order is matched, the parcel is in transit and the carrier gave an estimate |
| Return or exchange | The order date, delivery date, items, whether the return window is open, past returns from this customer | Labels the intent and the item; drafts the return instructions your policy gives | Approving the return, any exception to the window, the refund or the exchange | No |
| Damaged or wrong item | The order, the items shipped, the warehouse or 3PL that packed it, any photos attached | Labels the intent; summarizes what is wrong; drafts an apology that promises nothing | Replacement, refund or reshipment, and whether to ask for photos or a return | No |
| Address change | The order and whether it has been fulfilled or handed to the warehouse | Labels the intent; reads the new address from the message | Whether the address can still change, and making the change in Shopify or the 3PL | No |
| Wholesale or bulk enquiry | Whether the sender already buys from you as a trade customer | Labels the intent; extracts company, products and quantities | Trade prices, terms and whether to reply at all | No |
| Other or unclear | The customer's recent orders, if any | Labels it "other" and writes a one-line summary | Everything | No |
Order facts come from Shopify, not from the model. In Shopify's Admin API, a Fulfillment carries tracking information (the tracking company, number and URL), an estimated delivery date and the date it was delivered (checked October 3, 2026). If your orders ship from a 3PL that writes tracking back late, the workflow reads the 3PL's own feed too; a parcel with no tracking yet is never "in transit".
How does the confidence threshold decide who sends?
A model's label comes with a confidence score, and that score is the gate between a draft and a sent reply. The send rule below is strict on purpose. A reply may go out without a person only when every line is true. If any line fails, the reply waits as a draft in the assigned person's queue, with the reason it was held shown at the top.
- The intent is "where is my order", the only type on the auto-send list
- The model's confidence is at or above the threshold your team set, for example 0.90 to start (illustrative, not a recommendation for your data)
- The rules matched exactly one open order to the customer
- The fulfillment has tracking and is in transit, with a carrier estimate that is still in the future
- The tone is not negative, and the message does not mention a refund, chargeback, cancellation or complaint
- The draft contains only facts the rules fetched and text from an approved template
Set the threshold from your own tickets, not from a vendor's default. Run in draft-only mode for a few weeks and record, for every ticket, the model's label and the label your team would have used. The threshold sits where the model's high-confidence labels match your team's on almost every ticket. Helpdesks expose the same idea in their own way: Zendesk's intelligent triage gives each prediction a confidence field, and Shopify Inbox applies a topic label only when its system has high confidence (both checked October 3, 2026).
What does the workflow do, step by step?
The step map above is the whole workflow. Two steps use a model: labelling the ticket and drafting the reply. Everything between them is rules, and the send gate after them is a rule too. That order matters. The model reads messy text well, but it should not decide which order a customer meant or what a carrier said, so the lookup happens by rule against Shopify and the carrier, and the model only writes up what the rule found.
Routing is a rule as well. A damaged-item ticket goes to whoever handles replacements, an address change to whoever can still reach the warehouse before it ships, and a wholesale enquiry to the person who sets trade prices. A ticket with a negative tone or a high order value can jump the queue. Most brands already route this way by hand; the workflow only does it the moment the ticket arrives.
Which tool should run it?
Check your helpdesk first. The three common choices already label tickets, and the table sets out what each does on its own and where a separate workflow takes over.
| Level | Fits when | What to know |
|---|---|---|
| Gorgias intents and rules | One or more Shopify stores, support already in Gorgias | Gorgias detects intent and sentiment on incoming customer messages on all Helpdesk plans, with intents such as Shipping/Status, Shipping/Change, Order/Damaged and Return/Request, and leaves the field blank when it is not confident. Its auto-tag rule templates include one that tags shipping tickets differently by the order's Shopify status (checked October 3, 2026). It lists no wholesale intent, so that route needs a rule of your own |
| Zendesk intelligent triage | Support in Zendesk, often across several brands | Intelligent triage classifies topic, sentiment, language and entities on Suite and Support Professional plans and above; using the classifications in workflows requires the Copilot add-on (checked October 3, 2026) |
| Shopify Inbox | Chat on your store, small volume | Shopify Inbox is a free app from Shopify. It labels messages with topics including Order status, Edit orders, Returns/Delivered orders and Wholesale, only in English, and offers suggested replies that Shopify's guide says to review for tone and accuracy (checked October 3, 2026) |
| A workflow tool with an AI model step | One inbox and one store, or a helpdesk without built-in triage | Zapier, Make or n8n can call a model, look up the order and post a draft. Log every run, because a step that fails quietly leaves a ticket unanswered |
| Custom code on your accounts | Several stores or marketplaces, a 3PL's tracking feed, or a send gate your helpdesk rules cannot express | The lookup, the gate and the log live in one place you control, and the helpdesk stays the place your team works |
Permissions matter as much as the model. Gorgias's Shopify Actions let an agent cancel, refund or edit an order and change its shipping address straight from the ticket (checked October 3, 2026). That is useful for your team. For the automation, connect it with an account that can read orders and write drafts, and nothing more. When custom code is the right answer, it can run on Ecomsellertool Growth OS beside your other order flows.
What can go wrong, and how do you catch it?
The costly failures are confident ones: a reply that sounds right and is wrong. Each needs a check that fails loudly.
| What goes wrong | How it shows up | The check that catches it |
|---|---|---|
| The ticket is matched to the wrong order | A reply quoting someone else's tracking | Auto-send needs exactly one matched order; a customer with two open orders always gets a person |
| The model promises something | "We'll send a replacement today" in a draft | Templates hold every promise; a draft with text outside the template is held |
| Tracking is stale | "In transit" for a parcel stuck for days | The carrier estimate must be in the future; past-due parcels go to a person as late |
| An angry customer gets a cheerful template | A complaint that escalates | Negative tone, or words such as refund or chargeback, blocks auto-send |
| The label is wrong on a new kind of question | A run of edited drafts for one intent | The weekly sample compares labels with what your team changed; a rising edit rate lowers the auto-send list |
| A non-English message | Low confidence or a blank label | Low confidence always goes to a person; Shopify Inbox's topic labels are English-only |
Marketplace buyer messages run under each marketplace's own messaging rules, so they need their own templates and checks. Our guide to automating Amazon operations covers buyer messages separately. For deciding whether support is the right first project at all, see what to automate first.
Who builds this for ecommerce brands?
Ecomsellertool builds support triage like this on your own helpdesk and store accounts. We have built software on Amazon's seller APIs since 2017 and shipped 50+ tools for 100+ teams. We map your ticket types, templates and refund rules first, then build the lookup, the drafts and the send gate. Send a project brief below, or read how our workflow automation service works.
The pieces are work we have shipped:
- AI that classifies and drafts while a person sends. For Seller7 we built a review module that pulls in low-star reviews, classifies intent, suggests the next action and drafts a reply in the buyer's language, with brand-tone templates and human approval on every send. The team reviews, edits and sends.
- Carrier data in one place. For Ecom Global we built a shipping layer with carrier types, rate cards and routing rules, built for 100M+ products, 20+ marketplaces and 13+ carriers: the kind of data a where-is-my-order reply depends on.
Emailed purchase orders often land in the same inbox as support tickets; our B2B order entry recipe handles those.
Get this workflow built
Send us your version of this workflow
Describe the workflow you want automated. We reply with a written plan: scope, timeline and a fixed price.
Frequently asked questions
Can AI answer customer support tickets for my online store?
It can draft most of them and send a narrow few. A model is good at reading what the customer wants and writing a reply in your tone. It is not a safe judge of refunds, replacements or exceptions to your policy, so in this recipe it drafts those and a person decides. Only order-status replies built from live tracking data may go out on their own.
Which support tickets are safe to send without a person?
Where-is-my-order questions where the order is matched with certainty, the parcel is in transit and the carrier has given a delivery estimate. The reply only repeats facts from your systems. Returns, damaged items, address changes, cancellations and wholesale enquiries all involve a decision or money, so they stay as drafts for a person.
How do I set the confidence threshold for AI ticket triage?
Run the workflow in draft-only mode first and compare the model's labels with what your team chose on a few hundred tickets. Set the threshold where the model's high-confidence labels agree with your team on almost every ticket, then allow auto-send for one ticket type only. Keep sampling sent tickets every week and raise the bar if errors appear.
Does Gorgias or Zendesk already classify tickets with AI?
Yes. Gorgias detects intent and sentiment on incoming messages on all Helpdesk plans. Zendesk's intelligent triage predicts topic, sentiment, language and entities on Suite and Support Professional plans and above, and using those predictions in workflows requires the Copilot add-on. Shopify Inbox labels chat topics in English. Start with what your helpdesk already offers.
Can an AI tool issue refunds or cancel orders from the helpdesk?
Helpdesks such as Gorgias let an agent refund, cancel or edit an order from the ticket, so an automation connected with the same permissions could too. In this recipe it cannot: the automation's account gets no refund or cancel rights, and those actions stay with a person who sees the draft, the order and the customer's history.
Do I need custom code for AI support triage?
Not always. If you run one Shopify store on Gorgias or Zendesk, their built-in intent detection and rules may be enough. Custom code earns its cost when tickets have to match orders from several stores, marketplaces or a 3PL, or when you want a strict, logged send gate that your helpdesk's rules cannot express.
