Customer Support

Building an AI Customer Support Agent That Actually Works

How we built a support agent that resolves 73% of tickets autonomously — without hallucinating company policies.

Sergiu Poenaru·February 21, 2026

The Problem

A mid-size eCommerce company was drowning in 2,000+ support tickets per week. Their team of 8 agents couldn't keep up. Response times averaged 14 hours. Customer satisfaction was at 62%.

They'd tried a basic chatbot. It answered FAQs but escalated everything else — which was 80% of tickets.

The Agentic Workflow

We built a multi-step AI agent using Claude that could:

  1. Classify the ticket (refund, shipping, product question, complaint, other)
  2. Retrieve relevant context from their knowledge base (RAG over Zendesk articles + Shopify order data)
  3. Draft a response following their brand voice guidelines
  4. Execute actions (issue refunds under $50, update shipping addresses, cancel orders)
  5. Escalate to a human when confidence was low or the customer was upset

Key Design Decisions

Results (After 8 Weeks)

MetricBeforeAfter
Auto-resolution rate12%73%
Average response time14 hours2 minutes
Customer satisfaction62%89%
Tickets per agent/day4512 (complex only)

Tech Stack