E-commerce (under NDA)
AI customer-service agent that resolved most repetitive tickets and freed the human team for complex cases.
Problem
The customer received around 4,000 weekly inquiries across WhatsApp, email and web chat. Almost 70% were repeat questions about shipping status, return policies and product availability. The customer-ops team grew faster than sales.
Solution
We built an agent connected via RAG to their catalog (over 12,000 SKUs), policy documentation and order management events. The agent solves first line and escalates to human only on ambiguous intent or VIP accounts. Guardrails prevent refund promises outside policy.
Stack
- Claude 3.5 Sonnet
- Node.js
- Postgres
- Pinecone
- Vercel
"In four months we recovered the investment. The customer-ops team now focuses on what truly needs a human."
Context
The retailer ran three digital storefronts (own marketplace, Mercado Libre and WhatsApp Commerce) with a six-person customer-ops team across two shifts. The company was growing ~18% month-over-month in orders, but support cost was growing at the same pace: any bad hiring month pushed SLA above 6 hours.
Technical decisions
We used Claude 3.5 Sonnet as the main agent for its Spanish reasoning quality and stable tool use. Context is injected via RAG (Pinecone + re-ranking) over:
- A normalized SKU catalog with attributes.
- Internal FAQs and return/exchange policies.
- Order management events (state, tracking, history).
The agent uses typed tools (lookup_order, lookup_stock, escalate_to_human) with permission validation and usage logs. Per-user rate limits.
Lessons learned
The biggest initial risk was not the AI but catalog hygiene: inconsistent titles and empty metadata degraded answers. We invested two weeks in an enrichment pipeline before launching the agent, and that’s where the real unlock happened.