AI Customer Support Agent

A SaaS company's support team had gone from five people handling a manageable queue to five people drowning in 2,400 tickets a week. Average response time had drifted to 48 hours. CSAT had slid from 4.1 to 3.2 in six months. Customers were filing complaints on Twitter, which triggered more support tickets, which made everything worse. They needed an AI layer that could resolve tickets rather than acknowledge them with a canned response, and escalate cleanly when it couldn't.

We built an agent on top of the Claude API with RAG-based knowledge retrieval from the company's existing help docs, product specs, and past resolved tickets. The agent runs on Django with a React-based embeddable widget; WebSocket connections via Redis give the chat a real-time feel. When the agent hits something it can't resolve (billing disputes, account merges, legal escalations), it hands off to a human agent with the full conversation history, the relevant account context, and a suggested resolution path. Humans don't restart from zero. They pick up at 80%.

• 70% of incoming support tickets auto-resolved without human intervention
• Average response time for AI-handled tickets: under 2 minutes (was 48 hours)
• CSAT jumped from 3.2 to 4.6 within 90 days of deployment
• Support team headcount stayed flat through 3x customer growth the following year

Claude was chosen over GPT-4 for this use case because it follows complex support workflows more reliably and hallucinates less on edge cases. RAG uses embeddings stored in Redis for sub-100ms retrieval; vector DBs would have been overkill for a ~5K-document knowledge base. Django handles the conversation persistence and analytics; WebSockets keep the chat responsive. The escalation logic is its own module, and we tune it monthly based on which auto-resolved tickets get reopened.