avatarin has deployed a voice agent based on GPT-Realtime in Japanese Yamada Denki stores. Multilingual support, round the clock. In two weeks, 30,000 people talked to the agent and 92% of satisfaction survey responses were positive.

Voice was consolidating as an operational channel for enterprise agents, as we covered on July 24. This is the first published case where a voice agent holds up at real scale, not in a lab.

For anyone building with AI, the point isn’t the 92%. It’s that someone took a voice agent out of demo and ran it night and day. Voice is the hardest channel: no time to re-read, the customer won’t wait, and a two-second delay in response sounds like a failure. Seeing a real deployment helps you understand where we actually are, even if the announcement only tells the side that works.

What’s missing is the flip side: what happened with the 8% of negative responses, what does it cost to keep the service running, how do you handle conversations that go off the rails. The case is published by OpenAI, which has every incentive to show the win.

In detail

What came before. Retail voice systems were IVR: tree menus, «press 1 for…», limited keyword recognition. They worked because they were deterministic, but couldn’t understand a sentence outside the script. Then came chatbots with text-to-speech: they transcribed, generated text, converted back to voice. Three separate components, with cumulative latencies that made conversation feel unnatural.

What changes with GPT-Realtime. OpenAI’s Realtime API handles streaming audio: the model receives audio, processes it, and responds in voice in the same flow, without intermediate steps. For those not building it, the difference is between a system that «listens, thinks, speaks» as one thing and a system that glues three separate pieces together. Latency drops below the threshold where conversation becomes acceptable to a human.

The avatarin case. The agent provides multilingual support to Yamada Denki customers, an electronics chain. It’s not a general assistant: it answers questions about products, hours, availability. The scope is narrow, and that probably contributes to the 92% satisfaction. An agent that does few things has fewer chances to fail.

The numbers, with caution. 30,000 users in two weeks and 92% positive responses are the numbers from the case study published by OpenAI. We don’t know the survey methodology, sample size, or what counts as «positive». The 8% of negative responses goes unanalyzed. Without these details, the 92% is an indicator, not a measurement. It’s like reading the benchmark from whoever built the model: useful, but take it for what it is.

What stays open. We don’t have data on cost per conversation, abandonment rate, or edge cases where the agent had to hand off to a human operator. A real deployment is evaluated on these metrics, not just stated satisfaction. For anyone wanting to build something similar, the questions to ask are the same ones we cover in the lesson on costs, latency, and security: what’s the maximum latency your customer will accept, and what happens when the agent doesn’t know the answer?