Small sellers that adopted AI tools built into a Chinese e-commerce platform recorded higher monthly sales than matched non-adopters, according to a November 2026 working paper.
Focusing on firms with annual sales below 5 million yuan, the researchers estimated a 51.6% increase in monthly sales associated with adoption. The estimated number of orders rose by 47.0%, and the largest proportional gains appeared among the smallest firms.
The figures come from one platform and one set of firms. They are estimates from the researchers’ methods, not a guaranteed return for a business that buys an AI subscription. The National Bureau of Economic Research paper is a working paper and has not been presented as a final peer-reviewed result.
AI delivered through business software
The platform embedded generative AI in operating software used by sellers in 2024. The tools drew on centrally developed models and platform-wide data, giving firms access to applications without requiring each seller to build its own system.
Panle Jia Barwick, Lu Fang, Jinhan Liu, Huilan Xu and Zhe Yuan linked records of AI use to monthly operations and sales. They compared adopters with matched non-adopters over time and used additional statistical methods to address differences between firms.
A matched difference-in-differences design compares changes in outcomes for two groups before and after a development. It can help account for stable differences, but its conclusions depend on assumptions about how the groups would have changed without adoption.
More information and less routine work
The researchers report that adopters produced more product content, changed listings more often and spent less staff time per order. They also reached more customers and introduced products more quickly.
One proposed mechanism is a reduction in the cost of routine work. Another is that richer product information helps customers find and evaluate goods. The paper reports larger gains in settings where demand was harder to predict and matching customers with products was more difficult.
The results do not show that every extra listing improves a buyer’s experience. Product details still need to be accurate, and a faster content workflow can spread mistakes if descriptions are not checked.
Our coverage of an experiment with startups examined a separate question: whether examples helped firms identify useful tasks for AI.
Sales are not profit
A rise in sales is different from a rise in profit. Sellers must also consider advertising, fulfillment, returns, staffing and any fee for the software. A store with more orders may not earn proportionately more if the extra revenue is expensive to deliver.
The study’s setting also matters. Platform-level data and ready-to-use tools may help firms in ways that are unavailable to a retailer using a standalone service. The researchers’ estimated gains apply to the firms and period analyzed, not to all industries or countries.
For a small business, a sensible evaluation would compare accurate listings, completed orders and contribution after costs with a suitable baseline, while monitoring the time employees spend checking and correcting generated material.
