- 5×listing throughput
- -64%first response time
- 6languages supported
Starting point
After the client’s cross-border e-commerce platform had been running for a while, the bottleneck moved from technology to operations: six country sites, thousands of SKUs, and every new product needing titles, selling points and descriptions in six languages. Support handled a constant stream of repetitive enquiries. Tens of thousands of reviews went unread.
The client initially wanted an all-purpose assistant. We suggested cutting it to the three most labour-intensive scenarios and doing those properly.
One: multi-language listing generation
Not translation. Expression conventions, compliance wording and keyword preferences differ by market — for the same insulated bottle, the US site leads with capacity and heat retention while the German site cares more about material certification.
The approach: product attributes, selling-point material, market compliance rules and high-converting competitor phrasing form the context; generation runs per market and then passes a compliance term check against per-market prohibited-claim lists. Output enters a human confirmation queue, so operations edit and publish.
Throughput went from “two days for one product in six languages” to “a draft in half a day, publishable after edits.”
Two: review insight
Tens of thousands of multilingual reviews are structured by AI into issue type, component involved, sentiment intensity and repurchase impact. Aggregated weekly, it tells the product team things like “negative reviews about lid leakage on the German site tripled in the last fortnight.”
The value here is not technical — it turned data nobody had time to read into an input for product decisions, and drove two structural product changes within two months.
Three: support copilot
Deliberately not full automation, but agent assist: when support opens a ticket, order status, tracking, conversation history and relevant help articles are already retrieved, with a draft reply prepared for the agent to confirm or edit.
Two reasons for assist over automation: cross-border support involves returns policy and dispute handling where mistakes are costly, and the human confirmation step continuously produces high-quality training data.
First response time fell 64%. The support team was not reduced — their time moved to complex disputes and key accounts.
Outcomes
- Listing throughput up roughly 5× across six languages
- Review insight reports became a standing input to product iteration
- Support first response time down 64%, ticket backlog eliminated
- All three scenarios share one AI gateway: unified metering, caching, rate limiting and a cost dashboard
Related
The underlying commerce platform is described in Cross-Border E-Commerce Platform; the evaluation and cost method is part of AI Application Development.