关键要点
- AI drafts responses; humans always review and edit before sending—never auto-send
- Use ticket triage prompts to sort by priority, category, and escalation path
- Define confidence threshold (e.g., 0.8); below threshold, escalate to human agent
- Store customer context in prompt: Tier (VIP/standard), history, account status
- Monitor response quality: Track % helpful (customer feedback), CSAT, escalation rate
Support Workflow Stages
From ticket arrival to resolution: ingest → triage → draft → review → send → track.
- Stage 1 — Ingest: Ticket arrives; extract: subject, body, sender tier, account history
- Stage 2 — Triage: Prompt classifies: category (billing, technical, feature request), priority, escalation flag
- Stage 3 — Draft: Prompt generates response based on category + history + company tone
- Stage 4 — Review: Human agent reads draft; edits tone, adds personalization, approves
- Stage 5 — Send: Agent clicks "Send"; ticket moves to done or escalation queue
- Stage 6 — Track: CSAT survey sent; feedback loops back to improve prompts
Triage Prompt Design
Classify tickets in one pass; output: category, priority, escalation yes/no, and confidence.
- Input: Ticket subject + body + customer tier (VIP/standard) + account age
- Output: JSON with category (enum), priority (1–5), escalate (boolean), confidence (0–1), reasoning
- Categories: Billing, Technical, Feature request, Bug report, Account access, Other
- Escalation rules: Escalate if (a) confidence <0.7, (b) VIP tier, (c) Legal/security keywords
Response Template Prompts
Category-specific prompts; each one knows the context and tone for that type of issue.
- Billing prompt: "Customer says {issue}. Use friendly tone; offer specific next step (refund, invoice reissue, etc.). Keep under 200 words."
- Technical prompt: "Customer getting error {error}. Provide diagnostic steps. Assume customer is non-technical. Include links to docs."
- Feature request: "Customer requesting {feature}. Thank them; explain if on roadmap; offer workaround if available."
Include Customer Context
Personalization comes from context, not magic; feed history into prompt.
- Tier: "This is a VIP customer; response tone should be especially attentive"
- History: "Customer has contacted 5 times this month about {topic}; acknowledge pattern; offer escalation"
- Account status: "Customer on free trial; in response, mention upgrade benefits naturally (not hard-sell)"
- Previous conversations: Include last 2–3 interactions; model maintains continuity
Human Review is Non-Negotiable
AI drafts; humans send. No exceptions. Review includes tone, accuracy, and personalization.
- Review SLA: Agent must review draft within 5 minutes (for urgent tickets) or 30 min (standard)
- Edit options: (a) Send as-is, (b) Edit and send, (c) Reject and redraft, (d) Escalate
- Quality bar: If draft requires major rewrites >3 times, flag prompt to writers (sign of bad prompt)
- Metrics: Track % edited (ideal 20–30%; high % means prompt needs improvement)
Escalation Workflows
Define clear escalation paths; AI hands off below confidence threshold.
- Confidence threshold: If <0.7, escalate to human agent immediately
- Category escalations: All "billing refund" requests escalate to supervisor (compliance)
- Customer tier escalation: All VIP tickets escalate to senior agent regardless of category
- Sentiment escalation: If customer appears frustrated (tone analysis), escalate to empathy-trained agent
Monitor Support Metrics
Track quality from customer perspective; use feedback to improve prompts.
- CSAT (customer satisfaction): % rating response as helpful (target >85%)
- First-contact resolution: % tickets resolved without escalation (target >70%)
- Response time: Median time to draft + review (target <5 minutes)
- Escalation rate: % tickets escalated to human (should be <20%; rising rate = bad prompts)
- Agent editing: % of drafts edited; high % = prompt needs refinement
Common Mistakes
- Auto-sending AI responses without human review—high customer dissatisfaction, brand damage
- Generic tone ignoring customer tier—VIP customers frustrated by templated responses
- No escalation path—support agent forced to send inadequate response or escalate manually
- Ignoring customer history—repeating same solution customer already tried
- No feedback loop—prompts never improve because you're not tracking which are failing
Sources
- Intercom customer support AI practices, 2026
- Zendesk AI workflows guide
- Gorgias support automation case study