Avatarin Built a 24/7 Retail Agent on GPT-4 Realtime, Reaching 30K Shoppers in Two Weeks

Japan’s retail sector has long battled a structural constraint: experienced sales associates cannot staff every floor of a showroom 24 hours a day. Avatarin, a customer service AI company spun out of ANA Holdings, partnered with Yamada Denki, Japan’s largest electronics retailer, to solve that problem by encoding retailer expertise into a voice agent. According to OpenAI’s blog, the resulting Kurashi-Marugoto AI Agent, built on GPT-4 Realtime, reached 30,000 users in a two-week trial, with 92% reporting positive experiences.

How Avatarin Chose GPT-4 Realtime

The decision to build on OpenAI’s unified multimodal model reflects a shift in how retail AI should work. Avatarin CEO Akira Fukabori emphasized that the company needed a single system capable of handling speech, text, and visual inputs without the latency overhead of chaining specialized subsystems.

According to the OpenAI blog, Fukabori described the core advantage: “For us, the performance of a single model went beyond what we had seen from specialized speech recognition systems. It can work across speech, text, and images with low latency.” This unified architecture meant Avatarin could eliminate the conventional chatbot’s reliance on keyword matching, allowing the agent to respond to uncertainty, shifting customer preferences, and the contextual nuances that turn generic product suggestions into actionable recommendations.

Avatarin had prior experience with OpenAI’s API for speech recognition and inquiry analysis, but the GPT-4 Realtime integration represented a first opportunity to deploy those learnings directly to end consumers at scale.

Encoding Retail Expertise Into Conversation Design

The agent’s architecture reflects three design priorities. First, product information must remain accurate without introducing pauses that break conversational flow. Avatarin used retrieval-augmented generation (RAG) to ground the agent’s responses in Yamada Denki’s product database, ensuring current pricing and specifications remained available within GPT-4 Realtime’s reasoning loop.

Second, the company incorporated Yamada Denki’s sales methodology into the agent’s prompting and conversation flows. According to the OpenAI blog, customer service knowledge varies significantly by product category—a refrigerator recommendation requires different diagnostic questions than a washing machine purchase. By translating that category-specific expertise into conversation design, Avatarin transformed a generic Q&A system into a domain-adapted consultant.

Third, the system was designed to handle the nuanced scenarios that reveal the gap between chatbots and human service. Fukabori illustrated the difference with a concrete example: “Customers do not want a chatbot. They want intelligence. ‘I need a refrigerator for a family of four, but my kitchen is small. Which one should I choose?’ Real customer service means being able to answer that question.” GPT-4 Realtime’s capacity to track multiple constraints (household size, space constraints, budget) in a single conversational turn enabled this kind of multidimensional reasoning.

Why This Matters

The trial’s 92% satisfaction rate suggests that voice-first, context-aware agents may finally be displacing keyword-driven chatbots in customer-facing retail—a shift that has been predicted for years but rarely demonstrated at scale. For Yamada Denki and other retailers facing labor constraints, the implication is immediate: AI can extend expert availability beyond store hours without sacrificing the consultative quality that drives purchase confidence.

For model developers, the result vindicates the architectural bet on unified multimodal systems. If avatarin’s results hold up across different product categories and retail markets, GPT-4 Realtime’s latency advantages over chained pipelines will become a competitive table-stake in customer service applications. The next question is whether this pattern spreads beyond Japan’s appliance market to other retail verticals—automotive, furniture, financial services—where product complexity and customer uncertainty are equally high.