01

What changed

Most shopping AI stories ask whether the bot can answer more questions. This paper asked a better everyday question: what happens when a shopping app finally lets people use the chatbot in the language they actually want to use?

The researchers worked with a mobile-commerce platform in India that had spent almost two years offering only an English chatbot. For six days, new users were randomly assigned to either Version 1, where the chatbot stayed in English, or Version 2, where the chatbot could switch between Hindi and English in the same conversation. The firm did not change the rest of the app design or purchase flow. That made the language switch itself the key experimental difference.

The first result looked like a straightforward win. Across the raw, model-free comparisons, purchases per session rose 146% and chatbot interactions per session rose 132% in the bilingual version. But the paper did not stop at the optimistic read. It also found that uninstall rates were 107% higher in the bilingual version. In the regression analysis the authors attribute an 87% increase in purchase odds and a roughly 76% increase in uninstall hazard to the localization change.

The most useful part of the paper is where the gains and losses split apart. Simpler, lower-involvement transactions such as utility payments benefited from localization. More complex, higher-involvement transactions that demanded more search, comparison, and input created more churn. The same feature that made the app feel more accessible at the easy end could make the experience feel more frustrating once the task became complicated.

02

What this could change for you

For shoppers, the practical benefit is easy to understand. Buying or paying through an app can feel much less risky when the assistant speaks the language you naturally think in. That is especially true for routine transactions where the job is mostly to confirm details and move forward without confusion.

For app teams, the paper is a warning against simplistic localization wins. Adding a local language can increase engagement and purchases, but that does not mean the surrounding workflow is ready. If the payment steps, selection flow, or product navigation still feel awkward, the same bilingual feature can raise expectations that the rest of the product fails to meet.

The honest use case is narrow but valuable: local-language AI looks strongest when it removes friction from routine purchases, especially for people who might otherwise bounce because the app feels too English-first. The paper argues for designing the full path, not just the chatbot surface.

03

What it does not prove

This was one mobile-commerce platform in India, run over only six days. The result does not prove that the same mix of gains and churn would hold in grocery delivery, banking, ride-hailing, U.S. retail apps, or markets where the language mix and customer expectations are different.

The bilingual treatment changed only chatbot language, not the rest of the purchase flow. That makes the experiment clean, but it also means the paper does not test what would happen if checkout, search, payment help, and error recovery were redesigned for bilingual use at the same time.

The result is not "vernacular chatbots always work." The paper's own evidence says the opposite: they worked best in low-involvement transactions and appeared to create more user frustration in high-involvement transactions that demanded more steps and more cognitive effort.

The bottom line

AI did not solve mobile commerce in general. It solved a narrower and more believable problem: making routine shopping interactions feel more usable when the chatbot can switch into the shopper's language. In this randomized field experiment, that advantage was real, but so was the cost. The practical lesson is to localize the whole buying journey, not just the chatbot.

Primary research

Language translation effects in Chatbots: Evidence from a randomized field experiment on a mobile commerce platform

Journal of Business Research · 2025 · DOI 10.1016/j.jbusres.2024.115158

View the research ↗