Japan has a familiar problem: appliance stores can’t always offer expert help outside business hours. Have you run into that moment when you need advice but the store is closed? The solution? avatarin turned human sales knowledge into a 24/7 shopping agent that speaks by voice and text, and can understand images.
What avatarin did with GPT-Realtime
The project, called Kurashi-Marugoto AI Agent, launched with Yamada Denki (part of Yamada Holdings). In a two-week public campaign on the online store around 30,000 people tried it, and 92% of responses in the follow-up surveys were positive.
This agent isn’t a traditional chatbot that waits for keywords. It uses GPT-Realtime to keep conversations natural with low latency, and it combines voice, text, and vision into a single coherent experience.
The goal was to move a human salesperson’s expertise into a conversational interface that actually helps you decide what to buy.
How it works — explained without technical jargon
avatarin built the agent around three concrete priorities:
-
Keep product information accurate without slowing the conversation. They use a system like
retrieval-augmented generationto anchor answers to relevant product data, whileGPT-Realtimepreserves conversational flow. -
Translate Yamada Denki’s sales experience into conversational design. Each product category needs different questions: buying a fridge isn’t the same as choosing a TV. The agent learns to adapt the question flow and to return to the topic when you get sidetracked.
-
Ask questions instead of just answering. Rather than waiting for you to say exactly what you want, the agent probes: "Do you have a small kitchen? How many people live in your home?" That turns a passive interaction into guided discovery.
To understand why this matters: imagine you ask, "I need a fridge for a family of four, but my kitchen is small. Which one do you recommend?" A keyword-based chatbot could fail. This agent responds taking into account space, capacity, and budget—just like a human salesperson would.
"This is not an extension of the conventional chatbot. It’s the start of an interface revolution in retail," says Fukabori from avatarin.
Results and practical lessons
The experiment numbers are clear:
- ~30,000 users in two weeks.
- Multilingual 24/7 support by voice and text.
- 92% positive responses in the post-use survey.
But the most valuable part was the qualitative insight. Every conversation revealed why customers hesitate, what information they lack, and how their preferences change in real time. Plus, each interaction ended with a short voice survey, which made it easy to collect natural, continuous feedback.
Customers described the agent as "easier to talk to than a real salesperson" and appreciated not feeling pressured to buy. For the store, that means not just sales, but data to improve service, marketing, and the next experience design.
Impact for retail and where this is headed
avatarin imagines a future where the same intelligence accompanies you on the web, by phone, and in physical stores. One thread of context, one cohesive brand experience: "One Intelligence. One Brand. Every interface."
OpenAI helped structure complex prompts, optimize costs for an always-on voice service, and share implementation best practices. That let avatarin focus on what matters: making the conversation feel accurate, fast, and aligned with the brand.
It’s not about all companies sounding the same. It’s about building an intelligence that truly embodies each company’s identity while staying consistent across touchpoints.
Final reflection
Are we witnessing the end of the cold-response chatbot? Probably. What this experience brings is much closer to how people actually talk and decide: context, doubts, and changing opinions in real time. For retailers who want to support customers outside traditional hours, this is no longer science fiction—it’s a practical tool.