Case study · Ecommerce
RAG-powered support assistant for a multi-store retailer
A retrieval-grounded customer support assistant that knows the product catalog, the returns policy, the order pipeline and the brand voice, and hands off to a human the moment it isn't sure.
The challenge
What we had to solve.
Customer support was drowning. Three brands, two warehouses, one small team, and 60% of tickets were variants of the same six questions about sizing, shipping windows, returns and order status.
Earlier chatbot attempts had hallucinated policy answers and frustrated customers, so leadership was nervous about anything "AI", but the alternative was hiring four more agents.
Our approach
How we tackled it.
We built a retrieval-augmented assistant grounded in three sources: the product catalog, the policy library, and a live read of the order system. The model is never allowed to make up a policy, if it can't cite a source, it escalates.
Critical workflows (returns, address changes, order cancellations) go through guarded action templates, not free-text responses. The bot can do the thing, but only via approved tools.
A confidence threshold and a human handoff button live on every conversation. Below threshold? Straight to a human, with the conversation history attached.
We instrumented everything, answer source, confidence, customer satisfaction post-chat, and what the human did when handoff happened. That feedback loop is what made the system get better.
Deliverables
What we built.
Specific, named outputs, not vague "strategy".
RAG retrieval layerIndexes product catalog, policy library, FAQ archive and refreshes on edit.
Action toolkitReturns, address changes, order lookups and cancellations as approved tool calls.
Human handoffConfidence-gated escalation with full conversation context attached.
Analytics dashboardResolution rate, satisfaction, source citation rate and a backlog of common unanswered questions.
Voice + tone trainingBrand-aligned response templates, tested against a held-out set of real customer messages.
Outcomes
What it returned.
- 74% of tier-1 tickets resolved without a human touching them, verified against held-out test set.
- Customer satisfaction score on bot-handled tickets: 4.6 / 5, higher than the email-support baseline.
- The support team got their afternoons back to focus on complex cases, returns prevention and proactive outreach.
- Zero hallucinated policy answers in three months of production, because the bot will refuse rather than invent.
The takeaway
What we learned.
The thing that earned trust wasn't the model, it was the guardrails around the model. Let the AI do tier-1 confidently, escalate cleanly when it isn't sure, and measure everything. Customers prefer that to either bad bot answers or 6-hour email replies.
Services used
How we got it done.
Related work