AI Customer Service Handles 70% of Queries, CSAT Up 22%

The Situation

the client operates 200+ home goods stores across North America, with a growing e-commerce business that had tripled in volume over the previous two years. Their 12-person customer support team was stretched to breaking point — handling over 500 tickets per day, the vast majority of which were the same 40 questions asked on repeat.

Order status. Return windows. Store hours. Product availability. Discount code validity. These questions consumed most of the team's day, leaving complex customer issues — the ones that actually required human judgment — waiting hours or days for a response.

The Challenge

the client had already tried two chatbot solutions: a rule-based one that frustrated customers with its rigid decision trees, and a third-party AI tool that gave confident but frequently inaccurate answers about their specific policies and inventory.

The requirements for a successful solution were demanding:

• Real-time order status via Shopify API integration — not just canned responses
• Accurate policy answers grounded in the client's actual policy documents — no hallucinations
• Seamless handoff to a human agent when the chatbot couldn't help, with full conversation context passed to Zendesk
• A tone and persona that matched the client's brand — warm, helpful, not robotic
• Mobile-first UI embedded on their e-commerce site and accessible via a standalone link for in-store staff

Our Approach

The hallucination problem from their previous solution was the highest-priority risk to solve. We addressed it with a Retrieval-Augmented Generation (RAG) architecture: all factual answers are grounded in a vector database of the client's actual policy documents, product information, and FAQ content — not generated from the model's training data.

For real-time data (order status, store hours, inventory), we built direct Shopify API integrations that the chatbot queries on demand. The model only generates language — it doesn't make up facts.

The escalation path was designed carefully: the chatbot detects frustration signals (repeated questions, explicit requests for a human) and hands off proactively, passing the full conversation context to Zendesk so agents don't have to ask customers to repeat themselves.

The Solution

We delivered a production-ready AI support assistant with:

• RAG knowledge base: 340 indexed documents covering return policies, shipping windows, product care, store policies, and promotional rules — updated weekly via automated sync
• Live order lookup: Real-time Shopify order status, tracking links, and delivery window estimates — available to any customer who provides their email or order number
• Policy Q&A: Grounded, accurate answers to policy questions — the chatbot cites the relevant policy section and offers to send a copy via email
• Intelligent escalation: Automatic handoff to Zendesk with conversation summary, customer sentiment score, and relevant order/account context pre-populated for the agent
• Multi-surface deployment: Embedded widget on the client's e-commerce site, standalone URL for in-store staff, and a read-only analytics dashboard for the CX leadership team

Results

In the first month post-launch, the chatbot handled 70% of incoming queries without human intervention. Daily tickets reaching human agents dropped from 500+ to approximately 150 — letting the support team focus entirely on complex, relationship-critical interactions.

Counterintuitively, CSAT scores improved by 22%. The combination of instant response times, accurate answers, and smooth escalation outperformed the previous experience of waiting hours for a human agent to respond to a simple order status question. the client has since expanded the chatbot to their in-store kiosk network.