Corporate pitch decks for enterprise artificial intelligence routinely promise structural shifts in top-line growth. In retail and consumer platforms, vendors argue that natural language agents will reinvent product discovery, expand basket size, and transform passive browsers into high-intent buyers. The pitch assumes that generative interfaces expand consumer willingness to pay.

When a technology is sold on top-line expansion, the first empirical question is whether consumers actually spend more money or simply face less friction on the way to the checkout counter.

A research team from Columbia Business School, Zhejiang University, and an international e-commerce partner recently published the first large-scale causal evaluation of generative tools across consumer-facing operations. In a study titled “Generative Artificial Intelligence and Sales Productivity: Field Experiments in Online Retail” (forthcoming in Management Science, doi:10.1287/mnsc.2026.00623), Lu Fang, Zhe Yuan, Kaifu Zhang, Dante Donati, and Miklos Sarvary tracked millions of users across seven randomized field experiments conducted between September 2023 and June 2024.

The researchers evaluated seven operational workflows deployed by a global cross-border retail platform: pre-sale customer support chatbots, search query refinement, product description generation, push notification copy, ad title optimization, chargeback dispute defense, and live chat translation. Because all seven experiments ran within the same technical infrastructure and market conditions, the data isolates how generative models perform across different baseline environments.

Their findings provide a concrete benchmark for enterprise adoption. Generative deployments produced positive, measurable revenue gains in four customer-facing workflows. When aggregated across those applications, the annual incremental revenue generated by the technology was roughly five dollars per consumer (between $4.60 and $5.20).

For a platform serving tens of millions of active buyers, five dollars per user represents a meaningful aggregate sum. However, the operational value of that revenue depends entirely on how the gains were achieved and what it cost to generate them.

The study documents that the entire revenue lift came from higher conversion rates rather than increased spending intensity. Across every workflow with positive results, transaction frequency rose while average cart value remained completely flat. Consumers did not buy premium items, add extra products to their orders, or increase their overall budget allocations. They simply completed purchases they might otherwise have abandoned due to confusing product specifications or language barriers.

This distinction explains why the benefits were concentrated heavily among inexperienced consumers. First-time buyers on a cross-border platform often struggle with translated search terms and uncertain product details. For those users, automated intent expansion and multilingual descriptions reduced navigation friction. Experienced buyers, who already understood catalog navigation and product specifications, saw almost zero incremental benefit from generative assistance.

The data also reveals the steep diminishing returns of model intervention against existing software baselines.

The largest individual sales lift in the study occurred in the pre-sale customer service chatbot, which recorded a 16.3% increase in consumer expenditure. That gain occurred because the control condition offered zero live pre-sale support. The model was competing against an empty queue. Where the platform had already deployed mature systems, the marginal benefit dropped sharply. In search query refinement, where the control group used standard machine learning search with lexical translation, adding generative query expansion lifted sales by only 2.1%. In product description generation, the lift was 2.7%.

In creative marketing tasks, the generative advantage disappeared entirely. The researchers evaluated automated models against human copywriters for marketing push notifications and external advertising titles. Both experiments yielded zero statistically detectable impact on sales. Generating hundreds of algorithmic variations did not improve consumer response in crowded feeds where attention is already scarce.

These results define the true unit economics of consumer-facing language models. Generative interfaces function as digital lubricants. They reduce information and communication friction in messy catalogs, helping marginal buyers locate inventory. They do not manufacture consumer demand.

For engineering leads and product architects, this empirical reality sets a hard ceiling on infrastructure spending.

If a generative deployment produces five dollars in annual gross revenue per active user, the entire margin can be erased by inference bills and platform software licensing. In an e-commerce architecture where an active user performs dozens of catalog searches, reviews multiple product pages, and queries a conversational shopping assistant, the cumulative token volume accumulates rapidly. If those interactions rely on frontier model API calls or third-party enterprise vendor seats charging several dollars per seat or fractions of a cent per prompt expansion, the serving cost quickly outpaces the gross margin generated by the conversion bump.

An enterprise captures value from generative retail tools under two specific conditions. First, the application must target an absent baseline where simple automation replaces an empty workflow. Second, the technical team must run small, task-specific models on owned infrastructure at sub-cent unit costs.

When an architecture treats generative models as a magic multiplier for consumer demand, it risks paying enterprise inference rates for five dollars of friction reduction.