How Jomashop Automated Its Highest-Volume Support Calls

Jomashop's support line takes around 26,000 calls a month. When the team looked at what those calls were actually about, roughly 70 percent of combined call and ticket volume came down to three things: where is my order, my package never arrived, and please cancel this order. Cancellations alone accounted for about 17,000 tickets a month.

None of those are hard questions. They are just relentless, and each one took an agent four and a half to five minutes to work through. Meanwhile the questions that genuinely need a person, a customs hold on a high value watch, a fraud review, a backorder with no clear date, sat in the same queue behind them.

Key challenges included:

• Three high frequency, low complexity intents absorbing the majority of agent time, at four and a half to five minutes of handle time each
• Backorder and customs delay visibility concentrated in one or two people, so the answer depended on who was working
• A catalogue spanning accessories to six figure watches, where a single support policy cannot fit every order
• Only around a quarter of calls carrying a logged reason, leaving the team without the data to see what was driving volume or target a fix
• Seasonal peaks concentrating demand at exactly the moments when service expectations are highest

Jomashop partnered with NuPlay to deploy a voice and chat agent across its highest volume support intents. Rather than building one agent that tries to cover everything at a shallow level, the team built a supervisor agent that works out what the customer needs and hands the conversation to a specialist built for that job alone.

The supervisor greets the caller, identifies the intent, and verifies the customer once. From there it passes control to one of three sub agents: order tracking, cancellations, or lost and delayed orders. No sub agent re verifies the customer, and none of them hand off to each other, so the customer never repeats themselves and never gets bounced between flows.

Key workflows included:

• Intent identification and routing. The supervisor determines which of the three requests it is hearing and routes to the matching specialist. Anything outside that scope, returns, exchanges, damaged items, gift cards, price adjustments, is acknowledged and passed to a person rather than half answered.
• Verification built for a fraud sensitive catalogue. Identity checks were hardened through testing. Zip code was dropped entirely because too many customers share one. Phone number alone is never trusted, since it can be spoofed. Order ID sits as a secondary check and the name on the order is confirmed in every flow, with an alternate phone number available as a fallback path.
• Cancellation gated on live item status. Before the agent tells a customer an order is cancelled, it checks whether that item can still be cancelled. If it has already shipped, the agent says so plainly instead of making a promise the system cannot keep.
• Lost package handling that matches the evidence. For a package reported missing, the agent confirms the delivery address only when the carrier shows the item as delivered, and raises a shipping claim to the team when one is warranted.
• Transfers that land in the right place. When a person is needed, the agent quotes the current wait time, offers the customer the choice of waiting or submitting a ticket online, and routes to the correct queue. High value orders take a priority route, Spanish speaking customers reach a Spanish speaking agent, and sales enquiries go to the sales line.
• Context written back, not lost. Escalations carry an AI generated summary and transcript back to the sales order, so the person picking up the conversation starts with what already happened.

The design decision that matters most sits at the first branch. A request the agent is not built to handle gets named and passed on straight away, before any verification, rather than being partially answered. Customers are not made to work through an identity check only to be told the agent cannot help them.

What made it work

A deliberately narrow scope. Phase 1 covers three intents and nothing else. Email, returns and RMAs, order edits, address changes, Spanish language handling and outbound calling were all consciously left out. On a catalogue where the risk per order varies by four orders of magnitude, knowing which calls are safe to automate matters more than automating everything.

Design that respects order value. Cancel my order is a routine request at one end of the catalogue and a serious one at the other. Rather than applying one policy, the build checks live item status before acting, routes high value orders to a priority queue, and keeps fraud sensitive decisions with people.

Measurement agreed before the build began. Containment, CSAT, transfer accuracy, escalation correctness and response latency were each defined with a measurement method in the first two weeks, jointly between the two teams. Nothing was retrofitted to make the results look good.

A launch shaped around the contact centre's day. Calls stop reaching the agent about ten minutes before the centre closes, so no customer is handed to a queue that is about to shut. After hours coverage comes later, once the agent has a track record during staffed hours.

With the agent live across voice and chat, Jomashop's three highest volume intents are handled end to end without a person, and the calls that genuinely need human judgement reach one faster.

• 65%+ of calls resolved without a human
• 50% reduction in average handle time
• 100% of calls logged with the correct reason
• 3 intents live, across voice and chat
• 6 weeks from kickoff to go-live

The logging figure is the quiet one, and it may matter most over time. Before this, only about a quarter of calls carried a recorded reason, so the team was guessing at what drove its own volume. Every call the agent takes is logged with its intent by construction. Jomashop now sees the shape of its contact demand for the first time, which is what makes the next round of decisions possible.

For Jomashop's customers: answers on the first call without waiting for an agent, no repeating an order number across handoffs, no being told an order is cancelled when it has already shipped, and a Spanish speaking agent when one is needed.

For Jomashop's team: the repetitive three quarters of the queue handled elsewhere, escalations arriving with the order context and transcript already attached, and specialists spending their time on customs holds, fraud reviews and high value orders instead of reading tracking numbers aloud.