Study Finds Shared Generative AI Infrastructure Can Drive Durable Growth for Chinese E-Commerce Firms

Generative AI has repeatedly been shown to boost performance on discrete tasks, from writing marketing copy to handling customer questions. What has been harder to pin down is whether those improvements accumulate into durable business expansion—and which kinds of firms, if any, benefit most. A new study examines that question in a real-world setting where AI is not adopted piecemeal, but delivered as a shared layer of business infrastructure.

The researchers analyze a major Chinese e-commerce platform that, in 2024, built a generative AI system directly into the operating software merchants use to run their shops. Rather than requiring firms to develop their own models, the platform drew on centrally developed AI and platform-wide data to provide ready-to-use tools to every merchant. That design allows the study to move beyond surveys or self-reported usage: the authors have comprehensive logs of each firm’s AI activity linked to monthly operational metrics and sales outcomes.

To estimate impact, the paper tracks how adoption changes business capabilities—such as customer reach and the pace of product introductions—and how those shifts translate into revenue. The core analysis focuses on smaller merchants, defined here as firms with annual sales below RMB 5 million. Using a matched difference-in-differences approach, the authors find that firms adopting the embedded AI experience a large jump in performance: monthly sales rise by 51.6% after adoption. The increase is driven mainly by volume rather than price, with a 47.0% gain in the number of orders. The study also reports instrumental-variable estimates that corroborate the main results, addressing concerns that faster-growing firms might simply be more likely to adopt.

The operational footprint of the AI-enabled firms changes in ways consistent with genuine capability upgrades. After adoption, merchants reach more customers, refresh product lines more quickly, and improve product and service quality. Although the average gains are substantial, the proportional improvements are most pronounced among the smallest firms in the sample—suggesting the technology may narrow, rather than widen, gaps in performance within the small-business segment.

The authors point to two mechanisms behind these effects. First, AI appears to cut the cost of routine work. After adoption, merchants generate more product-related content and revise listings more frequently, yet spend less staff time per order. The benefits are largest for firms that, prior to adoption, had the highest operating costs relative to sales—consistent with AI functioning as a labor-saving process innovation where time was previously a binding constraint.

Second, AI seems to ease information frictions that impede matching between buyers and sellers. Adopters provide more detailed and richer product information, and the sales gains are strongest in markets where demand is difficult to forecast and where matching quality is poor. These effects are especially important for small firms, which often struggle to compete when product discovery and trust are major hurdles.

Importantly, the study argues that the impact is not limited to redistributing sales from non-adopters to adopters. At the market level, product categories characterized by higher pre-existing information frictions see growth along multiple margins: more sales by incumbent firms, more new entrants, and more new product introductions. In the authors’ telling, embedded AI expands markets rather than merely reallocating demand.

The paper notes that its conclusions are those of the authors. It also acknowledges feedback from seminar and conference participants and reports financial support from Chinese research funding programs for several co-authors, while clarifying that the views do not necessarily reflect those of the National Bureau of Economic Research.

Why it matters

If AI is delivered as a platform-wide utility rather than a bespoke tool, it may act less like a niche productivity add-on and more like infrastructure—potentially reshaping competition among small businesses. The evidence here suggests the biggest proportional gains can accrue to the smallest firms, especially in markets where information problems suppress trade. The next question to watch is whether these gains persist as AI becomes ubiquitous, and how platforms’ control over models and data affects who captures the long-run value.