多くの社内向けLLMは、空の入力欄から始まります。日本の非技術職の利用者には、Googleよりも Yahoo! JAPAN に近い見せ方が必要だと考えます。
There has been a lot of discussion over the years about how Japanese web design philosophy differs from what is popular in the West.
The Western ideal is sparse. Google’s search page is still a logo, a prompt box and a lot of white space. The design assumes the user knows what they want and asks them to type it.
Yahoo Japan is the classic counterexample. Its front page carries news, weather, shopping, finance and entertainment at once. Rakuten Ichiba and Kakaku.com are denser still, with product listings, point offers, badges and price comparisons on every screen. To a Western eye this looks cluttered. To many Japanese users it looks complete, because everything they might want to do is visible before they choose.
Edward Hall’s distinction between low-context and high-context communication is one way to read the difference. In a low-context culture the message is explicit and the listener is not expected to fill gaps, so a page that asks one question and waits for a clear answer makes sense. In a high-context culture, much of the meaning sits in the situation and the relationship. A dense page supplies that situation up front, and the user sees the whole field before deciding where to go. This is my interpretation, and it matches what I have seen working inside Japanese companies.
I think the same split applies to how companies put AI in front of their employees, and it explains something I watched happen.
In 2025, a global consumer goods company I worked for rolled out Microsoft Copilot across its regional operating units. In the Japan and South Korea unit, adoption trailed other regions. The licenses and training materials were the same everywhere, and so was the tool. What differed was the context the tool was being used in.
Most enterprise LLMs open on a white screen with a text box in the middle and a cursor blinking in it. It is the Google front page, moved into the office. If you are using a chatbot to plan a weekend trip, that works fine. You know what you want to ask, and a mediocre answer costs nothing.
Work is different. Most people already know the small hesitation of a blank Word document or an empty Excel grid. An enterprise LLM window is the same blank page, with the added problem that it is a new tool and nobody has told you what it is for. An employee has a role, a team, internal rules and a task they may not yet be able to describe in a sentence. Facing an empty box, they have to decide what the tool can do, what they are permitted to put into it, how to phrase the request, and whether the answer can be trusted. In that unit, many people looked at that list of decisions and went back to the tools they already knew. Accounts existed, and most of the activity came from a small group of enthusiasts.
AI transformation (Ax) has to go beyond giving employees a training manual and an enterprise LLM account. For non-technical users in Japan, I think it needs to look more like Yahoo Japan than like Google.
The first requirement is context. The use cases should be defined and mapped in advance, by department and by task, and placed in front of the user. A sales administrator should be able to see that the tool drafts quotation follow-ups, summarizes meeting notes against the customer record and checks an order form for missing fields. That is the full menu, laid out the way Rakuten lays out a category page. The employee picks from what is visible and does not have to invent a prompt.
The second requirement is handholding. An employee who is unsure about a new tool will often wait until someone has shown how it fits their work. Walkthroughs, named internal contacts and worked examples built from the company’s own documents carry more weight than a general course on prompting. The training should be attached to the task, so the person learns the tool while doing something they already needed to do.
The third requirement is purpose-built systems. A general chat window leaves integration to the user. Most of the efficiency gain comes when ML and AI sit inside the systems people already use, such as the expense tool, the ticketing system, the contract template or the approval workflow. The employee clicks a button labeled with a task they recognize, and the model runs behind it. Nobody needs to know what a model is. In a company where decisions pass through formal approval routes, an AI function that is already part of an approved system is also easier to adopt than a separate tool that each person has to justify.
The fourth requirement is a defined boundary. Employees need to know in plain terms which data they can use, which uses are approved and who to ask when a case is not covered. Ambiguity here is a larger obstacle than any technical limit, and it is a governance question that belongs in the plan from the start.
Taken together, these are the choices Yahoo Japan and Kakaku.com made years ago for shoppers: show the whole picture first and make the next step obvious. In my framework for Ax, built on the 10-20-70 split between algorithms, data, and people and process, this is the 70 percent. The model is the smaller part of the work. Most of the effort goes into showing the people who will use it what to do.