AI retail agent 24/7: the Yamada Denki case study with GPT-Realtime

The timeline: from zero to 30,000 interactions in fourteen days

The project stems from the collaboration between avatarin , a Japanese startup specializing in telepresence and physical agents, and OpenAI. The goal was precise: to eliminate downtime in the customer service of Yamada Denki stores. Yamada Denki is one of Japan's largest consumer electronics chains, with a diverse clientele in terms of language and age group.

Avatarin integrated GPT-Realtime — OpenAI's model optimized for low-latency voice conversations — within a physical agent deployed in stores. The result is documented directly by OpenAI in its official case library : 30,000 interactions in the first two weeks, with a 92% satisfaction rate among those who completed the post-interaction survey.

Plus, the system handles requests in multiple languages without manual setup for each one. This feature is key: Yamada Denki serves a growing number of tourists and foreign residents. A multilingual agent reduces reliance on specialized staff and slashes scaling costs.

Solution architecture: what makes GPT-Realtime suitable for physical retail

GPT-Realtime is built to handle streaming audio with lower latency than standard models. In a retail setting, every second of perceived wait time impacts the experience. Therefore, choosing a real-time model isn't just for looks: it's all about keeping the conversation flowing smoothly.

The avatar agent combines three technological layers. The first is the language model, which interprets the request and generates the response. The second is the physical presence layer — the robot or terminal — which gives the agent a visual identity. The third is the integration system with product data and shelf availability.

In fact, an in-store conversational agent without access to real inventory would be useless. The real technical complexity lies not in the language model, but in the integration with existing management systems. This is the knot that any retailer must face before evaluating a similar deployment.

To dive deeper into AI architectures applicable to the Italian context, the team at SHM Studio — AI services offers a preliminary assessment of the necessary integrations.

Winners and losers: who really profits from this model

The most obvious winner is the retailer. A 24/7 agent eliminates the cost of nighttime or holiday staff for basic info duties. Plus, it scales without major extra costs: the fiftieth simultaneous user costs just as much as the first.

However, there are less obvious losers. The first is the retailer who adopts technology without a clear customer journey strategy. A poorly configured agent, with generic responses or disconnected from real inventory, causes frustration rather than satisfaction. Yamada Denki's 92% is not an automatic figure: it's the result of careful design.

The second potential loser is unqualified sales staff. The AI agent handles standard informational requests. Consequently, the value of human staff shifts towards complex consulting and closing sales. Those who do not evolve their role risk progressive marginalization.

Finally, there is a third scenario: the competitor who stands still. In medium-complexity retail — electronics, furniture, DIY — whoever is first to implement a reliable AI agent builds a perception advantage that's tough to catch up to.

SHM Studio's take: real scalability or an isolated case?

The question Italian marketing managers are asking is a fair one: can a Japanese case study be replicated in Italy? The answer is nuanced.

On one hand, the underlying technology is identical. GPT-Realtime is available via OpenAI API for any developer or agency. Therefore, the technical barrier is not geographical. On the other hand, the context is different: Italian customers have specific relational expectations, and an agent perceived as cold or mechanical can damage the brand instead of supporting it.

We at SHM Studio we notice that success stories all have one thing in common: experience design comes before technology choices. First, you define what the agent needs to do, what questions it should handle, and the tone it should use. Only then do you pick the model and build the integration.

According to research by McKinsey on the economic potential of generative AI , retail is one of the sectors with the highest capturable value from conversational automation. However, capturing that value requires integration with existing processes, not just installing a chatbot.

Operational implications for Italian retail: three concrete scenarios

For marketing leads considering a similar project, it helps to break it down into three maturity levels.

• Basic level — digital informational agent: an AI assistant on the e-commerce site or app, powered by GPT-Realtime or similar models, that answers questions about availability, shipping, and return policies. Low investment, measurable impact in just a few weeks. Perfect for retailers with established digital traffic.
• Intermediate level — omnichannel agent: the same integrated agent between digital channel and physical store, with real-time inventory access. Requires integration with ERP or POS. The Yamada Denki case fits here.
• Advanced level — proactive agent: the system doesn't just answer requests, it anticipates needs based on browsing behavior or purchase history. This level requires a quality data foundation and a strategy of Digital marketing integrated.

For each level, the critical piece isn't the AI model, but the quality of product data and how well the brand voice stays consistent. An agent that doesn't match the retailer's vibe will just confuse the customer.

Metrics to track after go-live

Yamada Denki's 92% satisfaction rate is an output KPI. But process KPIs are just as important to understand if the agent is really working.

First of all, the containment rate : how many requests are resolved by the agent without human escalation. A rate below 60% indicates training or scope definition issues. Secondly, the average resolution time : it has to be lower than that of the human equivalent channel, or else the perceived value is zero.

Furthermore, it's essential to monitor the conversational drop-off rate : how many users drop off before getting an answer. A high rate points to perceived lag or irrelevant replies. Lastly, the post-interaction conversion rate : in retail settings, the agent must contribute to the sale, not just provide information.

To structure an effective measurement system, the google ads campaigns and the activities of SEO can be tied in with the agent's data to build a seamless funnel from acquisition to conversion.

What nobody is saying: the cultural context problem

The Yamada Denki case takes place in Japan. This is not a neutral detail. Japanese culture has a very high technological acceptance threshold, especially for robots and automated agents in public places. The average Japanese customer is used to interacting with automated systems in contexts where an Italian customer would expect a human.

Therefore, replicating the case in Italy requires a change management phase for the customer, not just for internal staff. Communication about the launch must be carefully handled. The agent should be presented as an additional service, not a replacement. The tone should be warm, not technocratic.

According to Harvard Business Review , the negative perception of AI agents in physical settings is often linked not to the technology itself, but to a lack of transparency about the system's automated nature. Openly stating that it is an AI agent increases trust, rather than decreasing it.

This is a principle that we at SHM Studio we also apply this to content design and digital communication: transparency is a competitive asset, not a weakness. For those who want to dive deeper into how to build an effective narrative around these initiatives, the service of SEO copywriting and that of LinkedIn campaigns offer practical tools to communicate innovation to stakeholders and B2B customers.

Next moves: what to consider in the next six months

The avatarin-Yamada Denki case represents a useful benchmark, not a model to copy. Every retailer has a different context: customer mix, catalog complexity, internal digital maturity.

Therefore, the first recommended step is an audit of the current customer experience: where are repetitive requests concentrated? Which questions receive slow or inconsistent answers? Which channels are uncovered during periods of low staff presence? These three questions identify the correct scope for a first AI agent.

Next, the prototyping phase should be quick and measurable: two to four weeks, one channel, a limited set of intents. Yamada Denki's time-to-value—two weeks—isn't a tech miracle. It's the result of a well-defined scope right from the start.

For anyone looking to dive into these opportunities with the help of a dedicated team, the first step is a consulting with SHM Studio . We support marketing teams from strategy definition to go-live, with a focus on the measurability of every phase. Additional resources and use cases are available in the SHM Studio blog and in the section web development for those evaluating deeper integrations with existing systems.