E-commerce customer-service agents: from cost centre to profit centre

The keyword for e-commerce customer service in 2026 is no longer cost cutting but taking orders. In figures published by Alibaba’s AI shop assistant Dianxiaomi, merchants that adopted it saw the hand-off rate to humans fall by 45% on average, and AI-plus-human collaboration converted more than 10% better than human-only service — the first time it has beaten humans on conversion.

Fix response time first

After one leading e-commerce platform adopted intelligent customer service, first-response time fell from 43 seconds to 0.8 seconds, and that alone cut potential order loss by 27%. Late-night shopping decisions are made in a very short window: wait a few minutes with no reply, the impulse fades, and the customer is gone.

From answering questions to making choices

JD’s Yanxiaozhi 5.0 was tested across 50,000 stores: hand-off to humans down more than 28%, satisfaction up more than 15%, pre-sales conversion up more than 37%; Samsung’s JD self-operated flagship store saw sales-advisory conversion rise 27.6%. On Taobao’s Dianxiaomi, Li-Ning’s official flagship store brought pre-sales hand-off from 40% to under 20%. The difference is that the latter systems actively call bundled recommendations, out-of-stock substitutes and precise payment reminders, rather than simply answering questions.

Retention and after-sales

In Dianxiaomi’s merchant test data, the average success rate for retaining an order against a refund is above 20% — recognising emotion and offering an alternative (a different size, a different model, a replacement shipment) costs less than going straight into the refund flow. High-frequency repetitive work such as after-sales triage and returns goes to the agent, and people move to emotion and disputes.

What it means for enterprises

• Before launch, work out three numbers: annual enquiry volume, average order value and margin. Without them, a conversion gain cannot be turned into money.
• Product and stock data must be clean first; dirty data makes the agent answer confidently wrong, which is worse than not answering at all.
• Keep an escalation path to humans. Boundary cases such as refunds and warranties cannot be guessed at.
• Feed conversation data back into product selection and ad targeting; do not use it only to count containment rate.