AI Support Triage Bot for n8n: Classify and Escalate

Most support queues are 60% repeat questions, 30% routine issues, and 10% genuine fires. The problem is that humans have to read all 100% to find the fires. This n8n workflow reads tickets first, answers the easy ones, and makes sure the fires hit Slack within seconds.

What this workflow does

• Webhook receives the ticket from your helpdesk (subject, sender email, message, ticket ID)
• OpenAI node classifies it: urgency (low/medium/high), category (billing, bug, how_to, account_access, feature_request, other), can_auto_answer, confidence, and — when auto-answerable — a complete suggested_reply
• Code node turns that into a route decision with conservative guardrails: anything high-urgency or low-confidence escalates; auto-replies only fire at confidence ≥ 0.8
• Switch node sends each ticket down one of three branches:
  • Auto-reply: HTTP request posts the AI-written answer back to your helpdesk API (as a pending reply, tagged ai-auto-reply, so agents can audit)
  • Escalate: formatted alert in #support-escalations with urgency, category, and a quoted excerpt
  • Log: routine tickets appended to a Google Sheets triage log

The guardrail logic is the important part. The LLM never auto-answers account-specific or low-confidence questions — the prompt explicitly restricts auto-answers to documentation-grade how-to and billing questions.

Prerequisites

• n8n instance reachable by your helpdesk’s outgoing webhooks
• OpenAI API key (gpt-4o-mini)
• A helpdesk with an inbound webhook and a reply API (Zendesk, Freshdesk, Help Scout, Intercom — the template uses a generic HTTP Request node you point at your API)
• Slack bot token; Google Sheets credentials for the log

How to import

1. Download the JSON above, then in n8n: Workflows → Import from File.
2. Attach credentials at the four placeholders (YOUR_OPENAI_CREDENTIAL, YOUR_HELPDESK_CREDENTIAL, YOUR_SLACK_CREDENTIAL, YOUR_GOOGLE_SHEETS_CREDENTIAL).
3. Point the HTTP Request node’s URL at your helpdesk’s reply endpoint and adapt the JSON body to its API shape.
4. Configure your helpdesk to POST new tickets to the production webhook URL, then activate.

What to customize

• Urgency definitions in the system prompt — the defaults treat security, data loss, and billing disputes as high; add your own SLA triggers
• The auto-answer whitelist — paste your top 10 FAQ answers into the system prompt so replies cite real product facts instead of generic guidance
• Confidence thresholds in the Code node (0.8 to auto-reply, below 0.6 escalates)
• Categories — rename them to match your helpdesk’s tag taxonomy so reporting lines up

For the full build, including a knowledge-base lookup step before the reply, see the AI customer support bot tutorial.

Cost per run

One classification call per ticket, typically 500-1,200 tokens round trip with gpt-4o-mini — as of mid-2026, comfortably under half a cent per ticket. A 2,000-ticket month costs a few dollars, which is hard to argue with when each auto-answered ticket saves several minutes of agent time.