Rakuten is giving every shop an AI manager. Nobody asked.

Every shop on Rakuten Ichiba is to get an AI avatar of its own store manager: available around the clock, handling product questions, recommendations and the walk to checkout. Announced alongside it was a "super agent" meant to reach across shopping, travel and finance, and to carry you all the way through booking and payment.

As technology, this is a large announcement. The most-upvoted comment underneath it was neither excitement nor fear. It was, roughly, users don't actually want this.

What I want to observe is not whether Japanese consumers approve of AI. It is how much trust the companies distributing it have already spent.

Context for readers outside Japan

Two things make this story read differently in Japan than it might elsewhere.

First, Rakuten is not Amazon. It is a mall: tens of thousands of independent merchants renting storefronts inside one platform, each with their own listings, their own photographs and their own reputation. So an "AI store manager" is not one company automating its own service desk. It is the landlord speaking in the voice of every tenant, to that tenant's customers. Who owns what the AI learns is a live question, not a philosophical one.

Second, Yahoo! News Japan's comment section is the closest thing the country has to a public thermometer: one thread per article, ranked by votes, no follower graph, no algorithmic feed. People who never post on social media vote there. This note counts 114 comments collected under one story on the day after the announcement.

One more piece of background matters, and it is the one Japanese engineers brought up immediately — see below.

What happened

On 5 August 2026, at Rakuten's "Rakuten AI Optimism" event, chairman and CEO Hiroshi Mikitani announced the AI store manager. Each merchant's manager and staff would be turned into an AI avatar carrying that shop's product knowledge and service style, on duty 24 hours a day, 365 days a year. A limited trial with some shops is planned before the end of the year.

The larger idea is the "super agent": a single AI front door across shopping, travel and finance, able to go as far as reservations and payment — the whole Rakuten ecosystem funnelled through one conversation.

Now the part the article does not say. It does not say what model the AI store manager runs on. It is described only as making use of "external AI."

That gap has a history. On 17 March 2026, Rakuten released Rakuten AI 3.0, presented as Japan's largest high-performance model and developed under GENIAC, a Japanese government-backed AI programme. Within hours of release, developers reading the model's config.json found the architecture declared as DeepseekV3ForCausalLM and the model type as deepseek_v3. The 671-billion-parameter mixture-of-experts design, with 37 billion parameters active per token, matches China's DeepSeek V3. Reports also noted that the original MIT licence file had been removed. Asked about the base model, Rakuten declined to disclose it.

Here is why that stings rather than shocks. Rakuten used to be the company that did say. RakutenAI-7B, released in 2024, was documented in a paper as continued pre-training on Mistral-7B — the attribution was explicit and unremarkable. The same company, on the same kind of question, stopped answering. The AI store manager arrives four months later.

What the comment section is saying

Start with the top comment.

84 upvotes, the highest in the 114 collected.

Notice what this comment does not do: it never questions whether the AI will work. Not accuracy, not capability, not jobs. It asks one thing only — who is this being installed for? People shop online to reduce friction. If service is being added back in, the suspicion is that it serves the seller.

This is not resistance to technology. It is suspicion of motive. And in Japanese comment sections, that particular suspicion travels furthest.

The next comment tests the suspicion instead of just voicing it.

Five upvotes. Almost nobody saw it. It is also the most rigorous thing in the thread.

This person did not stop at doubting. They went to the company's own AI, asked it whether advertising spend moves recommendations, got it to say yes, and brought the answer back. And then the conclusion: well, obviously.

That last part is what stayed with me. Someone verified a corporate claim, succeeded, and felt nothing. Distrust has become the baseline rather than the finding. You do not celebrate confirming what you already assumed.

And that distrust did not begin with this announcement.

78 upvotes, second place. "This isn't a conversation about AI at all" is the spine of the whole thread.

The same shape repeats above it. Third place (60 upvotes) asks not for AI but for the genuine cheapest price including shipping to simply be displayed. Fourth place (57 upvotes) asks not for AI but for more human beings answering the phone. Three of the top four comments are not discussing AI.

That is not incomprehension. It is an argument about order of operations. When the pictures don't match the goods, the search doesn't work and the phone doesn't get answered, adding a conversational layer on top is not an upgrade — it is a distraction. Reading it as ignorance would be the mistake.

The ones installing the AI become the ones it is installed on

There is a second constituency here that outside readers will miss: the merchants.

32 upvotes, and a level deeper than the consumer argument.

The AI store manager is a mechanism for moving a shop's knowledge into the platform. Once moved, whose knowledge is it? If a merchant leaves, does the trained manager leave with them? A tool that binds you more tightly the better you train it is not service software. It is lock-in.

The training cost also falls on the merchant. One commenter running chatbots professionally put the figure at about a year of continuous teaching to reach a usable standard, and asked who exactly, in an understaffed shop, is supposed to do that work.

Then there is this, from someone selling on the platform.

One upvote. Essentially nobody.

And yet this is the best-sourced comment in the thread. The first half is verifiable in the model's own configuration files, as described above. The second half is first-hand experience of the seller tooling.

So: a comment combining a checkable fact with direct experience gets one vote, while a comment closer to a feeling gets 84. That is worth stating plainly rather than pretending otherwise. Comment sections do not rank by correctness. They rank by how self-evident something feels. Anyone reading Japan through them — including me — has to hold both numbers at once.

What I make of it

Having read all 114, one thing is clear. These people are not afraid of AI.

"AI will take our jobs" barely appears near the top. What appears instead are call centres that loop forever, phones that never connect, and clothes that arrive looking nothing like the photograph. They are not reacting to artificial intelligence. They are reacting to what happened the last time humans were withdrawn and something automated was left in their place.

Then there is the sharpest line in the thread.

39 upvotes. Are you using the AI, or is the AI using you. That question is being asked by ordinary shoppers, in a thread about a shopping mall. This is what I would most like readers abroad to see: Japan's AI debate is not being conducted at conference altitude. It is being conducted at the height of daily life — can I return this, will someone pick up the phone, who is liable when it is wrong.

Here is my own reading.

What companies are shipping right now is frequently not an AI feature but the consumption of AI as a story. A tool that is genuinely useful when reached for deliberately gets stitched into an entire service so the firm can demonstrate — to investors, to staff, to competitors — that it is Doing AI. When that happens, the user stops being the beneficiary of a function and becomes set dressing for a narrative. Eighty-four people upvoted "nobody wants this" because being made into set dressing is easy to feel.

This is where Rakuten AI 3.0 comes back in. That the model is DeepSeek V3 underneath is not, to me, the offence. Building on an existing open model is a legitimate engineering decision, and Rakuten itself once documented exactly that with Mistral. The damage is not the base model. It is that a company which used to disclose stopped disclosing — announcing "Japan's largest," declining to name the foundation, and being identified from a configuration file within hours anyway. That sequence is what spends trust.

And it is the same company now proposing to stand an AI in every shopfront, wearing the shopkeeper's face, and to extend an agent across travel and money. When someone asks what that one is built on, is the answer undisclosed again?

Because an AI store manager is not the merchant's representative. It is the mall talking to the customer while wearing the merchant's face. A system like that has to be able to say whose interest it optimises for. If ad spend moves the ranking, publish that. If it cannot be published, that is the answer.

One last thing. The 78 people who said this isn't a conversation about AI at all should not be dismissed as laggards. They are not refusing the future. They are asking for promises to be kept in order. Ship what the photograph shows. Make search work. Answer the phone. Do those, and the AI store manager becomes convenient. Skip them, and it becomes a new entrance to the call queue that never connects.

The limited trial runs before the end of the year. I will be counting.

All quotes are verbatim, with sources cited and commenter handles withheld. Source articles may expire. The substance of this piece is the author's commentary.