Customer-support agent with Claude tool-use

Problem

Support team of 4 was drowning in tier-1 tickets — KB lookups, order status, simple refunds — leaving no bandwidth for hard issues.

Users

Public — embedded in the customer help widget, ~400 conversations/day

Key flow

1. 1Customer types question in widget
2. 2Agent picks tool(s), confirms intent if ambiguous, acts
3. 3Resolution or handoff — transcript logged for QA review

What I built

• Claude function-calling orchestrator with 6 typed TypeScript tools
• Vector KB search over docs + past resolved tickets
• Refund tool gated by order-age and amount thresholds
• Live handoff to human agent when confidence drops below 0.6

Limits

• No payment-method changes — always escalates to a human
• Knowledge cutoff: KB index is rebuilt nightly, not real-time