How to Use AI to Auto-Classify and Route Support Tickets
Support ticket triage is one of the most repetitive tasks in any customer-facing business — and one of the highest-ROI AI automation targets. Here is how to build a system that classifies, prioritises, and routes every ticket automatically.
The Cost of Slow and Inconsistent Triage
Poor triage does not just slow resolution — it directly damages customer relationships.
Response Time
Customers measure support quality primarily by first-response time. Every minute a ticket sits in an unsorted inbox is a minute of avoidable customer frustration. AI triage reduces first-response time from hours to minutes.
Routing Accuracy
A billing issue routed to a technical agent wastes both the agent’s time and the customer’s patience. Consistent AI classification ensures every ticket reaches the right person on the first attempt.
Data and Visibility
Manual triage produces inconsistent tags that make support analytics unreliable. AI classification with a fixed taxonomy produces clean data — enabling you to identify your top issue types, track trends, and measure the impact of product improvements on support volume.
The Step That Determines System Quality
Spend more time on your taxonomy than on your code. It is the foundation of everything else.
Audit 3 months of historical tickets
Export your last 3 months of support tickets to a spreadsheet. Read through 200 tickets and group them into natural categories. Most businesses find 8-15 categories cover 90%+ of volume. Keep categories mutually exclusive — a ticket should clearly belong to one category.
Define categories with examples
For each category, write a clear definition and 3 example tickets. This becomes your AI classification prompt context. Ambiguous categories produce inconsistent classification — if you cannot write a clear definition, split the category or merge it with another.
Add urgency levels
Define 4 urgency levels: Critical (service down, data loss, security issue — requires response in under 1 hour), High (core feature broken, blocking customer work — 4 hours), Medium (non-blocking issue or question — 24 hours), Low (feature request, general enquiry — 72 hours).
Map categories to teams and SLAs
For each category, define: which team handles it, which team lead is notified for high/critical urgency, and what the SLA target is. This routing map is what Make uses to direct tickets after classification.
Make + OpenAI Setup
// System prompt for ticket classification
You classify customer support tickets for a SaaS company.
Return ONLY a JSON object — no explanation, no preamble.
Categories:
- billing: payment issues, invoice queries, subscription changes, refunds
- technical_bug: product not working as expected, error messages, crashes
- account_access: login issues, password reset, 2FA problems, locked accounts
- feature_request: suggestions for new features or improvements
- how_to: questions about how to use existing features
- data_issue: missing data, sync problems, import/export issues
- performance: slow loading, timeouts, latency issues
- integration: third-party connection issues, API problems, webhook failures
Urgency levels:
- critical: service completely down or data loss risk
- high: core feature broken, blocking the customer
- medium: issue exists but workaround available
- low: question, feedback, or non-blocking request
Return: {"category": string, "urgency": string, "summary": string (one sentence),
"sentiment": "frustrated|neutral|positive", "suggested_tag": string}
📌 Include 2-3 example tickets and their correct classifications in the prompt. Few-shot examples improve classification accuracy by 15-25% for edge cases.
What Happens After Classification
| Category | Urgency | Routed To | SLA Target | Additional Action |
|---|---|---|---|---|
| technical_bug | critical | Senior engineer on-call | 1 hour | Page via PagerDuty, notify CTO |
| technical_bug | high | Engineering queue | 4 hours | Slack alert to #engineering |
| billing | any | Finance/accounts team | 4-8 hours | Flag if amount > threshold for manager |
| account_access | any | Support Tier 1 | 2 hours | Auto-check if password reset email was sent |
| how_to | any | AI auto-response | Immediate | AI drafts answer from knowledge base; human reviews |
| feature_request | any | Product team inbox | 72 hours | Auto-tag in product backlog tool |
| performance | high/critical | DevOps queue | 2 hours | Check status page; auto-update if known issue |
Closing Low-Touch Tickets Automatically
How-to questions are typically 25-35% of support volume. AI can resolve most of them without human involvement.
Detect how-to classification
When Make classifies a ticket as category: how_to and urgency: low or medium, branch the scenario to the auto-response workflow.
Retrieve relevant knowledge base articles
Use semantic search (OpenAI embeddings) to find the 2-3 most relevant help articles from your knowledge base. Pass the customer query and the retrieved article content to Claude.
Generate a personalised response
Prompt: ‘The customer asked: [query]. Relevant documentation: [articles]. Write a helpful, specific answer to their question using only the provided documentation. Address them by first name. If the documentation does not contain a clear answer, say so and offer to escalate.’
Send and monitor
Send the AI-generated response and mark the ticket as ‘AI-responded’. Monitor the customer’s reply: if they confirm resolved, close the ticket. If they reply with continued frustration or the issue persists, escalate to Tier 1 support immediately.
What classification accuracy can we realistically expect?
With a well-designed taxonomy and good examples in the prompt, expect 88-95% accuracy on first pass. The remaining 5-12% will be ambiguous tickets where human review is appropriate anyway. Track misclassifications weekly and add them as examples to your prompt to continuously improve.
What if a ticket could belong to multiple categories?
Design your taxonomy to be mutually exclusive, then add a secondary_category field to your JSON schema. For the primary routing decision, use only the primary category. Log secondary categories for analytics — they often reveal tickets that need a new dedicated category.
Should we tell customers their ticket was AI-classified?
You do not need to disclose AI classification — it is an internal routing mechanism, not a customer-facing AI interaction. The customer experience is: they submit a ticket, it reaches the right team fast, and they receive a quick, accurate response. That is the outcome they care about.
Want an AI Ticket Triage System Built for Your Business?
SA Solutions builds custom support automation systems — from ticket classification through auto-response — integrated with your existing helpdesk platform.
