AI Ticket Triage: A Complete Implementation Guide
End-to-end blueprint for building an AI triage layer in front of your help desk — schema, prompt, confidence thresholds and the human handoff.
Who this is for: Support leads and ops engineers who want to add AI triage to an existing help desk (Zendesk, Help Scout, Front, Intercom or HubSpot).
AI triage is the highest-ROI support automation: it's bounded, measurable and never replaces the human reply. This guide is the exact playbook Opsacea uses, including the JSON schema and prompt that survive real-world tickets.
What 'triage' actually means
Triage = classify + tag + route + (optionally) prioritize. It does not write the reply. Keeping the scope this narrow is what makes the project ship in 2 weeks instead of 6 months.
The output schema
- category: billing | technical | account | refund | sales | other
- urgency: low | normal | high | critical
- sentiment: positive | neutral | negative
- language: ISO-639-1 code
- suggested_owner: team key from your config
- confidence: 0.0–1.0 (the model's own self-rating)
Confidence thresholds
Below 0.65, do not auto-route — leave it for a human triager. Between 0.65 and 0.85, route with a 'low-confidence' flag visible to the agent. Above 0.85, route silently. Re-tune monthly using your false-positive log.
The handoff
- Write category/urgency/sentiment as tags on the ticket — never overwrite the agent's tags.
- Post the AI's reasoning as an internal note, never as a public reply.
- Log every classification + the agent's eventual correction to a sheet or DB for monthly review.
Measuring it
- Triage accuracy (agent correction rate) — target <10% after 4 weeks.
- First response time — target -50% vs. baseline.
- Mis-routed ticket rate — target <5%.
- Cost per ticket — should be sub-cent on modern models.
Next step
The Opsacea Toolkit ships the n8n workflow, the prompt file and the schema for this exact pattern — ready to wire into your help desk.
Keep reading
More from the same cluster — internal links that compound.