The Challenge
Japan's home-appliance retailers keep running into the same wall: shoppers want expert advice, but staffing is tight and stores close. According to OpenAI's case study, avatarin — an AI customer service company spun out of ANA Holdings — set out to solve this with Yamada Holdings. The goal was to take the knowledge of experienced sales associates and make it available around the clock, in more than one language, without piling extra work onto human staff. The result was the Kurashi-Marugoto AI Agent, a voice-based shopping assistant that guides customers from browsing to a purchase decision.
The Solution
The agent was built on OpenAI's GPT-Realtime. avatarin's CEO Akira Fukabori explained that the model's ability to work across speech, text, and images with low latency was the deciding factor. Instead of forcing people to phrase questions in a fixed way, the agent listens for context and responds naturally.
avatarin designed the system around three ideas:
- Accurate product information without slowing the chat. A retrieval-augmented generation (RAG) system grounds answers in real product data, while GPT-Realtime keeps replies quick.
- Sales expertise turned into conversation design. Since a sales associate needs different information for a fridge than for a laptop, Yamada Denki's customer service knowledge was built into the conversation flows and prompts, with guardrails to keep discussions on track.
- An agent that asks, not just answers. Rather than waiting for instructions, it asks follow-up questions to work out what a shopper actually needs.
OpenAI worked alongside avatarin to structure complex prompts, manage API costs for a voice service that runs continuously, and share implementation practices.
A Real Example
Fukabori gave a plain example of the problem the agent solves: "I need a refrigerator for a family of four, but my kitchen is small. Which one should I choose?" A keyword-based chatbot struggles with that. The avatarin agent instead asks about household size, available space, and budget, then recommends options grounded in Yamada Denki's product data. If you wanted to copy this pattern, your prompt design might instruct the agent: "Before recommending, confirm the customer's household size, room dimensions, and budget. Ask one question at a time. Only recommend products found in the connected catalogue." That combination of proactive questioning plus RAG grounding is the practical core of the blueprint.
The Results
In a two-week public campaign on Yamada Denki's online store, roughly 30,000 people used the agent, and 92% of survey responses were positive. The service offered 24/7 multilingual support by voice and text. Just as valuable were the insights: every conversation showed what shoppers cared about and where they hesitated. Customers could ask questions after closing time and talk openly about budgets without feeling pressure. Some described the agent as "easier to talk to than an actual sales associate." Each chat ended with a short voice survey, turning feedback into part of the experience.
Key Takeaways for SMEs
You do not need Yamada Denki's scale to apply the lessons. Ground answers in your own product data, encode your best staff's questioning habits into the conversation, and let the agent ask before it answers. avatarin frames its wider goal as "One Intelligence. One Brand. Every interface" — one assistant that carries a customer's context across web, phone, and store.
Brain.mt can help you use AI for your business. Contact me for more information. I also offer dedicated workshops and training about this subject.
