The problem
The platform had grown faster than its support infrastructure. With 200,000 active customers, the support team was handling 2,000+ tickets per week — mostly repetitive queries about order status, returns, and product compatibility. Average response time had ballooned to 48 hours. CSAT had dropped from 4.2 to 3.6.
The instinct was to hire more agents. We proposed a different model.
What we built
We built a support agent that combines three retrieval sources:
Live order data: Direct integration with the order management system allows the agent to answer “where is my order?”, “when will this be refunded?”, and “was my return received?” with accurate, real-time answers.
Product knowledge base: A RAG index over the full product catalogue, specification sheets, and compatibility guides. The agent can answer detailed product questions that previously required a specialist.
Resolved ticket history: 18 months of past tickets and resolutions, embedded and indexed. This gives the agent institutional knowledge that no new hire could replicate.
The escalation layer
This was the critical design decision. The agent is not trying to resolve everything — it’s trying to resolve the right things. We built a confidence-based escalation system:
- High confidence + routine query → auto-resolve and close
- Medium confidence → draft a response, flag for human review before sending
- Low confidence or complex issue → route directly to agent with full context pre-populated
Human agents spend zero time on routine queries and arrive at complex ones fully briefed.
Results
- 67% of tickets fully auto-resolved (no human touch)
- Average response time: 48 hours → under 4 minutes for auto-resolved tickets
- CSAT recovered to 4.4 within 8 weeks of launch
- Support team capacity freed to handle 3x the inbound volume without additional headcount