On July 30, OpenAI published a customer story about the AI shopping agent that ran on Yamada Denki's online store[1]. It was built by avatarin, a startup spun out of ANA Holdings, on GPT-Realtime, a single model that handles voice, text, and images. During a two-week public campaign, roughly 30,000 people used the agent and 92 percent of survey responses were positive[1].

An agent trained on how Yamada Denki associates actually sell

On a home-appliance sales floor, an experienced associate's explanation is often what closes the sale. Staffing is limited, though, and that expertise stops when the store closes. What avatarin and Yamada Holdings set out to do was move that knowledge into an agent that runs 24 hours a day[1].

The two companies signed a partnership in May 2024 to create AI and robotics services specifically for the home-appliance retail industry[2]. Building on that, they announced on February 25, 2026 that development had started on the Kurashi-Marugoto AI Agent. It uses a-commerce, avatarin's customer service and sales platform for online retailers, as its foundation, with the goal of covering everything from pre-purchase to post-purchase and daily life[3].

A defining feature is that the agent learns the way Yamada Denki staff talk to customers. Multimodal AI that understands text and voice together reads not only the words but the context and the situation, which allows the agent to guide shoppers from product consultation to a purchase decision around the clock, in multiple languages, by voice or text[3].

Roughly 30,000 users in two weeks, with 92 percent positive responses

The public campaign ran from 10:00 on March 3 to 12:00 on March 16 on a dedicated page inside Yamada Web.com, Yamada Denki's official online store[3]. Shoppers spoke to the agent from a PC, tablet, or phone, and instead of only answering questions, it worked out what they needed, made suggestions ahead of the request, and guided them toward a product[3].

According to OpenAI, approximately 30,000 people used the agent over those two weeks, and 92 percent of post-use survey responses were positive[1]. Each conversation ended with a short voice survey, turning feedback into a natural continuation of the shopping experience[1].

Akira Fukabori, representative director and CEO of avatarin, frames this as more than an extension of the conventional chatbot. He describes it as an attempt to expand what customer service can be with AI, and as the beginning of an interface revolution in retail[1].

One caveat is noted in the announcement: this effort is a lab focused on product selection through conversation with the agent, and it is not connected to information such as the Yamada membership app[3]. Linking it to the group's customer data platform is left to a later stage[3].

Designed to ask questions, not just answer them

avatarin says it designed the agent around three requirements[1].

The first is keeping product information accurate without slowing the conversation. A retrieval-augmented generation (RAG) system grounds answers in product data, while GPT-Realtime keeps responses fast[1].

The second is translating sales expertise into conversation design. The information an associate needs to gather varies significantly by product category, so Yamada Denki's customer service knowledge was built into the agent's conversation flows and prompting. The system follows along when a shopper changes requirements or drifts off topic, while guardrails keep the conversation anchored to the shopping experience[1].

The third is a policy of asking rather than only answering. A conventional chatbot waits for the next instruction; this agent asks follow-up questions to uncover what the shopper actually needs, moving from question-and-answer toward guided product discovery[1]. Fukabori illustrates the point with an everyday case: a customer who needs a refrigerator for a family of four but has a small kitchen. Being able to answer that, he says, is what real customer service means[1].

On why GPT-Realtime was chosen, avatarin points to a single model performing beyond what it had seen from specialized speech recognition systems, and to its ability to work across speech, text, and images with low latency. Rather than asking people to adapt to the rules of a machine, the technology can adapt to the way people communicate[1].

OpenAI helped with prompt structure and API costs

OpenAI worked with avatarin on structuring complex prompts, optimizing API costs for an always-on voice service, and sharing implementation best practices[1]. Because a voice agent stays live including idle time, cost design determines whether it is practical to operate.

Shoppers described the agent as easier to talk to than an actual sales associate, and said they appreciated being able to ask the same question again without stress[1]. The service also opened new moments for retail: customers could ask questions after stores closed and speak candidly about budgets or hesitation without feeling sales pressure[1].

avatarin also emphasizes that the conversation itself becomes data. What shoppers cared about, where they hesitated, and what might have helped them decide all surface in the dialogue, exposing information that was hard to see in conventional online shopping[1].

Toward one agent across web, phone, and stores

avatarin envisions the same intelligence carrying a customer's context across the web, the phone, and physical stores. Instead of separate systems for each channel, the experience stays continuous as one customer, one conversation, and one brand[1].

For Yamada Holdings, whose businesses extend from home appliances into housing and financial services, that vision leaves room for AI to support everyday life more broadly. At the same time, avatarin does not want every company's agent to sound alike. It wants people to experience an intelligence that embodies each company's identity and brand across every interface[1].

The agent was also demonstrated at Retail Tech JAPAN 2026, the 42nd trade show for distribution information systems, held at Tokyo Big Sight from March 3 to 6[3]. The next question is how much of the experiment can move into a permanent service.

Summary

The AI shopping agent that ran on Yamada Denki's online store drew roughly 30,000 users in two weeks, with 92 percent of survey responses positive. Its structure combines low-latency voice conversation through GPT-Realtime, product data grounded by RAG, and conversation design carried over from how store associates sell. For now the scope is limited to product selection as a lab effort, but the stated plan is to extend it into a foundation for customer service across channels.

Source: https://openai.com/index/avatarin/

Source: https://about.avatarin.com/info-news/news-release/8706/

Source: https://about.avatarin.com/info-news/news-release/9863/