There is a particular kind of shop assistant everyone has met. They appear at your shoulder within eleven seconds of you entering, ask if you're alright there, and then hover at a distance of roughly one and a half metres while you pretend to read the back of a box. You do not want their help. You want them to leave. You will later buy the thing online, at midnight, in your dressing gown, having read forty reviews.

Which makes the result out of Japan last week genuinely surprising.

On August 3, PYMNTS reported on a deployment at Yamada Denki, the Japanese home-appliance retailer, which turned its associates' sales expertise into a 24/7 multilingual voice agent built on OpenAI's GPT-Realtime. During a two-week public campaign on Yamada Denki's online store, about 30,000 people used the agent. Ninety-two percent of post-conversation survey responses were positive.

And in the feedback, customers said the agent was "easier to talk to than an actual sales associate."

The Agent Asked the Questions. That's the Whole Trick.

The agent was built by avatarin, an AI customer service company. Its CEO, Akira Fukabori, was blunt about what they were not building: "This is not an extension of the conventional chatbot. It is an attempt to expand what customer service can be with AI. We see it as the beginning of an 'interface revolution' in retail."

The structural difference, as PYMNTS describes it, is direction of travel. A conventional chatbot sits there waiting for you to type the magic keyword. It answers. avatarin's agent listens for context and asks follow-up questions to work out what you actually need — the way a trained associate would.

Fukabori's example is the one worth writing down: "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."

That is not a lookup. There is no keyword in that sentence that maps to a product code. Answering it requires holding two constraints in your head at once and knowing the range well enough to trade one off against the other. Under the hood, a retrieval-augmented generation system grounds the agent's answers in real product data, while GPT-Realtime handles voice, text and image understanding inside a single low-latency conversation — which is what lets the agent keep up when a shopper changes their mind halfway through.

Why People Preferred It to a Human

The satisfaction number is nice. The reasons behind it are more useful.

Per OpenAI's account, customers said they appreciated being able to ask questions again without feeling stressed. The format also let them speak candidly about budgets or uncertainty without feeling sales pressure — and it was particularly valued after stores had closed for the day.

Read that list again, because it is not a list of things AI is good at. It is a list of things humans are bad at. A person cannot un-hear that you asked the same question three times. A person is standing right there when you admit your actual budget. A person goes home at six.

The agent won on patience, on the absence of judgement, and on being awake. None of those are model capabilities. They're structural advantages that any always-on system has over a tired human on a Saturday afternoon.

The Customer Was Already Ready. The Business Usually Isn't.

If you suspect your customers would find this creepy, the wider numbers suggest otherwise. PYMNTS Intelligence research, conducted with Visa Acceptance Solutions across the U.S., Brazil and the UAE, found that 48% of online shoppers used AI to research their most recent purchase. Separately, PYMNTS Intelligence found 45% of consumers would be comfortable letting an AI agent complete a purchase on their behalf — rising to 54% among Gen Z.

On the other side of the counter, 43% of retailers are actively piloting autonomous AI shopping agents, and 81% say they already trust AI to operate autonomously once the right guardrails are in place.

Roughly half your customers are already doing this. The gate is no longer permission. It's whether you've built anything for them to talk to.

What to Actually Do With This

Here's the small-business translation, and it's cheaper than it sounds.

You do not need Yamada Denki's product catalogue. You need to notice that the winning behaviour in this story was asking, not answering. Most AI phone agents sold to small businesses are built as answering machines with better grammar: caller states need, agent retrieves fact, call ends. The Yamada Denki agent made money by doing the opposite — treating the first thing the customer said as incomplete, and digging.

So write down the three questions your best salesperson always asks before they recommend anything. Not the FAQ. The qualifying questions — how many people is this for, when do you need it, what went wrong last time. Then ask any vendor demoing you a phone agent whether it can ask those, unprompted, and change its recommendation based on the answers.

If it can only answer, you've bought a fancier voicemail. If it can ask, you've bought the bit of your best employee that goes home at six.