What changed
Livestream shopping mixes entertainment with a real buying decision. A host is showing products in real time while viewers are trying to decide whether a specific item is right for them. The paper starts from a simple problem: one human streamer cannot answer every shopper's questions quickly enough, so people buy with partial information or leave without buying at all.
The intervention was an AI streaming assistant that handled chat-based support during livestream sessions. In the randomized field experiment, shoppers in the treatment condition could use the assistant and shoppers in the control condition could not. The headline results were concrete commercial outcomes, not survey intentions: sales increased by 3.00%, and product return rates fell by 12.55%.
That combination matters more than a conversion lift alone. Many shopping tools can boost purchases by making people buy faster and regret later. This result points in a narrower and more useful direction. The paper argues that the assistant increased perceived intelligent information provision, which reduced decision uncertainty and raised confidence. The authors also report a tradeoff: AI interruptions could raise uncertainty, but the information benefit outweighed that friction overall.
What this could change for you
For shoppers, the believable use case is not an AI that shops for you end to end. It is an assistant that helps answer ordinary last-minute questions before you buy: size, fit, usage, compatibility, ingredients, or whether a product actually matches what the livestream host just claimed. If those answers are better and faster, fewer purchases should turn into avoidable returns.
For merchants and platforms, the useful lesson is that sales should not be the only success metric. The stronger signal in this paper is the pairing of a modest sales gain with fewer returns. That suggests a better shopping assistant does more than increase pressure. It can help buyers make a choice they are less likely to undo later.
For the broader consumer-AI market, this is one of the cleaner low-risk cases for automation in everyday commerce. The assistant did not need to replace the seller, set prices, or make a high-stakes judgment. It handled information friction inside a narrow workflow where a wrong answer is usually reversible and a good answer can save both buyer regret and merchant costs.
What it does not prove
This was one leading livestream selling platform in one shopping format. The paper does not show that the same effect will appear in ordinary search-based ecommerce, in-store retail, subscription renewals, or categories where products are more expensive, regulated, or difficult to compare.
The outcome gain was meaningful but not magical. A 3.00% sales lift is not a revolution, and the paper does not show that every product category or every session improved equally. The return-rate analysis also focuses on ordered sessions during the experiment, with a reported regression sample of 35,164, so the cleanest claim is about that observed slice of real purchases rather than every possible shopper interaction.
The paper's mechanism evidence is also narrower than the headline. The authors distinguish proactive and reactive AI interactions, but those interaction-mode analyses are correlational rather than the core randomized contrast. The study shows that an assistant can help in this setting. It does not prove that more autonomous shopping agents will always feel trustworthy, avoid interruptions, or improve long-term customer satisfaction.
The bottom line
This is a strong routine-commerce result. In a real shopping platform experiment, an AI assistant did not just nudge more people to buy. It also reduced product returns, which is a much harder metric to fake. The honest takeaway is that AI looks most useful here when it acts like a fast product explainer inside a narrow buying moment, not when it pretends to replace human judgment about what you should want.
Primary research
Artificial Intelligence (AI) Assistant in Online Shopping: A Randomized Field Experiment on a Livestream Selling Platform
Information Systems Research · 2025 · DOI 10.1287/isre.2023.0103
