How a global enterprise automated customer support with AI.

To answer one support ticket, an agent opened up to six tools. Zendesk held the ticket, Kibo the order, Qapla the shipment, and internal databases the rest. Most of the handling time went on gathering data before anyone typed a reply.

The client is an e-commerce partner to more than 50 luxury brands in fashion, design, and luxury goods. As those brands grew, so did the queue: order tracking, returns, address changes, and policy questions, in several languages.

The work before the answer

Agents sorted every ticket by hand and wrote every reply from scratch, including the ones they wrote every day. Returns were the slowest. Each one was checked against the return policy by an agent, one ticket at a time.

Response times depended on who picked up the ticket. The repetitive tickets filled the day, and the difficult ones waited.

What we found one level down

The brief was about automation. We started with the people doing the work: the support team, the operations managers, and the IT leads. Together we mapped where an agent's time went on a typical ticket, which tickets repeated, and where each piece of data lived.

The answer to most tickets already existed. It was spread across four systems that did not talk to each other. AI on top of that setup would have produced faster replies from incomplete data.

So we split the work into two phases. We brought the data together first, because a drafted reply is only as good as the order details behind it.

Phase 1: one screen for every ticket

We built a support console in Retool that brings Zendesk, Kibo, Qapla, and the internal databases into one view. An agent opens a ticket and sees the customer's order details and shipping status without switching tools. We built it so the AI step could later plug into the same data.

Phase 2: AI on the predictable tickets

Many requests follow a pattern: where is my order, I want a refund, please change my address. For these, AI categorizes the ticket, pulls out the order details, and drafts a reply in the customer's language. The agent reads the draft and approves it with one click.

Tickets that do not fit a pattern, including complex return and policy questions, are flagged for a person instead. The AI drafts, and an agent decides what is sent.

What changed

• Average handle time went from 3 to 4 minutes per ticket to 30 seconds.

• AI now handles 80% of incoming tickets, categorizing each one and drafting the reply for an agent to approve.

• More than 20,000 tickets have been resolved through the console.

Customers get the same quality of answer in their own language, whichever agent picks up the ticket. Agents spend their time on the tickets that need judgment. Every reply still goes out under an agent's approval. That is why the team could hand most of the queue to AI and keep control of the answers.