Customer support agent

Customer support agents field user questions, retrieve relevant articles, answer with citations, escalate when the question is outside scope. Simple concept; many failure modes in production.

The flow

1. User asks question
2. Agent retrieves relevant KB articles
3. Agent answers with citations
4. If question is out of scope: escalate to human
5. If question contains PII or policy content: apply guardrails
6. Log for review

Core tools

• search_kb(query), the main retrieval tool
• get_user_context(user_id), user history, plan, tickets
• create_ticket(summary, priority), escalation
• send_reply(user_id, text), actually respond

Key design choices

Retrieval quality is everything

See the RAG section of the site. Bad retrieval = agent has no good context = bad answers. Invest in hybrid retrieval, reranking, and eval.

Escalation must be fast and clean

When agent can't answer, don't fake competence. Escalate with all context preserved so the human doesn't start from zero.

Personalization with caution

Pulling user context helps answers but leaks can embarrass. Never cross user boundaries. Audit carefully.

Guardrails

• Don't promise refunds, discounts, or policy exceptions
• Don't acknowledge or deny account specifics you can't verify
• Escalate legal, threat, or abuse situations

Measurement

• Deflection rate (tickets handled without escalation)
• Customer satisfaction on agent responses
• Accuracy (did the answer match the right KB article?)
• Escalation rate
• False positive escalations (should have handled)

The trap: too eager to answer

Agent should say "I don't know, let me get a human" more readily than "here's my best guess." In support, confident wrong answers are worse than acknowledged gaps.