System Design: AI Customer Support with Tools, Escalation, and Quality Loops

AI customer support design interviews blend RAG, agents, and workflow automation. Interviewers ask how you retrieve policies and account context, when the bot can take actions versus suggest steps, and how humans inherit full context on escalation.

Ticket and CRM integration introduces freshness and permission challenges. The system must know open incidents, prior resolutions, and entitlements without leaking one customer's data to another. Candidates design context assembly pipelines with strict scoping.

Action tools—refunds, password resets, order modifications—require risk tiers and approval flows. Discuss idempotency keys, simulation modes, and rollback for partial automations. A support bot that double-refunds is worse than one that escalates.

Quality measurement goes beyond automation rate. Track containment, CSAT, repeat contact rate, and policy violation incidents. Interviewers want closed-loop learning from agent corrections without training on unreviewed chats.

Latency expectations in support are tighter than generic chatbots. Explain caching of stable policy snippets, parallel retrieval and account lookups, and streaming responses that show empathy while tools run. Operational playbooks cover peak events like outages when traffic 10x overnight.