An AI customer-service assistant resolving 72% of tickets without humans — support costs down 58%.
Where it started
RetailPro's support inbox grew 3x in a year as the chain expanded online. Response times stretched past six hours, reviews suffered, and hiring enough agents would have erased the e-commerce margin.
Previous chatbot attempts had failed badly — rigid decision trees that customers hated. Leadership was skeptical AI could handle real order, refund, and product questions without embarrassing the brand.
What we built
We built a retrieval-grounded AI assistant connected to RetailPro's live order system, product catalog, and policy documents — it answers with real order status and accurate policy, not generic responses. Confidence thresholds route anything uncertain to human agents with full conversation context.
A six-week pilot on 20% of traffic proved the numbers before full rollout. Systematic evaluation suites now test every prompt change against hundreds of real conversation scenarios before deployment.
- Retrieval-grounded answers from live orders, catalog, and policies
- WhatsApp and web chat channels in three languages
- Confidence-based human handoff with full context
- Evaluation suite of 400+ real scenarios gating every change
- Monthly ROI dashboard: deflection, CSAT, and cost per conversation
Drag to compare
- 12-second first response, 24/7
- Costs down 58% at higher volume
- 4.6/5 satisfaction on AI chats
- Three languages, every channel
Inside the build
“72% of our support conversations are handled flawlessly by the AI. Costs dropped by more than half while satisfaction went up. Remarkable engineering.”