Every store inbox contains tickets of very different weight. A customer asking about opening hours, another asking where their order is, a third saying their card was charged twice, a fourth threatening a dispute and a fifth reporting an allergic reaction. Treated in arrival order, the urgent ones wait behind the trivial ones, and the agent spends the first minutes of each ticket working out what it is about.
Triage fixes this. It is the first step in the support process: decide what the ticket is, how urgent it is and where it should go. Done well, it saves time, reduces errors and improves the customer experience. This guide offers a practical design built around rules, with AI helping where wording varies. It complements support macros, chatbot handoff rules and where is my order.
Step 1: Define categories
Fewer categories work better. Aim for eight to twelve that map to how your team works, each with a clear owner or macro set.
| Category | Examples | Typical handler |
|---|---|---|
| Order status | Where is my order, tracking, delivery date | Bot or macro |
| Order change or cancel | Wrong address, add an item, cancel before shipping | Orders team, time-critical |
| Returns and refunds | Return request, refund status, exchange | Returns team |
| Problem with item | Damaged, wrong, faulty, missing parts | Support, with photo handling |
| Payment and billing | Double charge, declined, invoice | Finance-aware agent |
| Product question | Size, compatibility, materials, stock | Product knowledge agent or bot |
| Discount and promotion | Code not working, price match | Support, with rules |
| Account and website | Login, checkout error, site bug | Tech support |
| Wholesale and partnerships | Business enquiries, press | Sales or owner |
| Complaint or escalation | Threats, dispute warnings, public complaints | Senior agent |
| Spam and other | Marketing solicitations | Auto-close |
Step 2: Set priorities
Priority should reflect what a delay costs, in money, risk or relationship, not only arrival time. A small set of levels is enough.
| Priority | Criteria | Target first response |
|---|---|---|
| Urgent | Safety issue, fraud or dispute threat, payment error, delivery within 24 hours of a change request, high-value customer upset | Within an hour in working hours |
| High | Order issues before shipping, damaged or wrong items, repeated contact on the same issue | Within four hours |
| Normal | Most questions and requests | Within one working day |
| Low | General enquiries, feedback without a problem | Within two working days |
Targets are examples; set yours according to your team and customer expectations, and publish them. Tickets that wait beyond target should escalate automatically.
Step 3: Detect with rules
Rules are fast, transparent and cheap. Start with signals that need no interpretation.
| Signal | Rule | Result |
|---|---|---|
| Order number present | Look up the order | Attach status; likely order status or item issue |
| Sender matches an order | Match email to orders | Attach recent orders and customer history |
| Keywords | “tracking”, “where is”, “refund”, “broken”, “chargeback”, “lawyer” | Category and priority hints |
| Order status and age | Order unfulfilled for longer than promised | Raise priority; add delay macro |
| Customer value | Many orders or high lifetime value | Raise priority |
| Channel | Marketplace messages, social, email, chat | Route by channel rules |
| Attachments | Photos present | Likely item problem; route to the right queue |
| Language | Detected language | Route to a speaker |
Step 4: Use AI for the messy rest
Real messages are vague, long and emotional. “Hi, I ordered something last week and it still has not turned up and my sister’s birthday is on Saturday” is an order status ticket with a deadline. A language model can classify it, extract the order reference or the deadline and summarise it in a line. Use AI as a classifier with a fixed set of categories and priorities, and require it to return structured output with a confidence score. Route high-confidence results automatically, send low-confidence ones to a human triage queue and log everything. Apply the principles from rules vs AI.
- Give the model the category definitions and examples.
- Provide only the data it needs; strip sensitive details where possible.
- Never let it take actions on its own such as refunds; it classifies and drafts.
- Detect sentiment and urgency cues, such as threats or distress, and raise priority.
- Sample outputs weekly and compare with agent corrections.
Step 5: Attach context
A ticket that arrives with the order, its status, tracking, the customer’s history and the relevant policy saves several minutes per ticket. Pull this data automatically on arrival and show it in the agent’s view. Include a one-line AI summary where useful. Agents should never have to open the store admin to answer the most common questions. Our customer support and order care and customer support chatbot services build these integrations.
Step 6: Route and respond
| Route | Condition | Action |
|---|---|---|
| Auto-reply | Order status with tracking available, FAQ questions with clear answers | Send a macro-based reply, open the ticket for follow-up if the customer replies |
| Draft for approval | Returns, simple refunds within rules | Prepare a reply and action; agent confirms |
| Specialist queue | Payment, tech, wholesale | Assign to the group |
| Escalation | Threats, safety, public complaints, VIP | Notify a senior person |
| Auto-close | Spam, no-reply messages | Close with a tag |
Be careful with auto-replies. A reply that misses the point is worse than a slower, correct one. Start with the safest categories, such as order status where data is certain, and widen as accuracy proves itself.
Handling special ticket types
- Threats of chargeback or legal action. Mark urgent, collect order and delivery evidence immediately and involve a senior person. See dispute evidence.
- Safety and health reports. Escalate at once, preserve details and follow your product safety procedure.
- Repeat contacts. A third message about the same order is a sign the first two answers failed; raise priority and assign to the same agent.
- Social media complaints. Respond quickly in public with a route to a private channel, and open a ticket to track.
- Fraud suspicion. Route to the fraud queue without accusing the customer; see order fraud rules.
- Data requests. Requests to access or delete personal data have legal deadlines; tag and route to the person responsible for privacy.
- Messages in other languages. Detect language, route to a speaker or use translation with review.
A short document listing these cases, the tag each gets and the owner prevents a lot of ad hoc decisions. Review it every quarter against the tickets that caused problems in the previous period, adding rules where the same mistake occurred twice. Training new agents is easier too, because the triage logic is written down and visible in the helpdesk itself.
Service level targets and escalation
- Set a first-response and a resolution target for each priority.
- Measure from creation, in working hours.
- Warn agents when a ticket is approaching its limit.
- Escalate automatically when a limit is passed.
- Review breaches weekly to find causes: staffing, routing or ticket type.
Measure and improve
| Metric | What it shows |
|---|---|
| First response time by priority | Whether triage puts urgent items first |
| Misroute rate | Tickets moved to another queue; a measure of accuracy |
| Agent tag corrections | Where rules or AI get it wrong |
| Auto-resolved share | Volume handled without a person, with satisfaction scores |
| Reopen rate | Poor or premature answers |
| Tickets per 100 orders | Overall friction; see the KPI dashboard |
| Top contact reasons | What to fix upstream |
What to do this week
If you want to try this in a small helpdesk without building anything, start with three rules. First, tag any ticket that contains an order number or the word tracking as order status, and attach the order. Second, raise the priority of tickets that mention a chargeback, a lawyer or a safety problem. Third, send spam and no-reply messages to an auto-close tag. Review the tags for a week, correct the mistakes by hand and note which ones came up. Those corrections become your next rules, and after a month you will have a triage system fitted to your own customers rather than a generic one.
A worked example
A store receives 3,000 tickets a month. Without triage, an agent spends about 45 seconds per ticket reading and categorising, plus 60 seconds looking up the order, around 87 hours. With rules and AI classifying 85 percent correctly and attaching order data, that falls to about 15 seconds of review, saving roughly 70 hours a month, or $1,500 at $22 an hour. Urgent tickets reach an agent within an hour instead of the next morning, and complaints about delayed responses fall. The classifier’s errors, about 15 percent of tickets, are corrected by agents, and those corrections are used to improve the rules. The figures are illustrative.
Mistakes to avoid
- Too many categories for anyone to remember.
- Triage by arrival time only.
- Auto-replies that do not read the message.
- No feedback loop from agent corrections.
- Sending customer data to tools without checking privacy terms.
- Never reviewing the rules.
Sources and further reading
Figures and rules in this guide come from the primary pages below. Platform limits and prices change, so open the page before you build on a number.
- Zendesk: triggers and automations, rules for tagging and routing tickets
- Gorgias Help Center: rules, ecommerce helpdesk rules
- Help Scout Docs: workflows, automating ticket handling