AI adoption in a small or mid-sized Japanese business usually stalls for reasons that have nothing to do with the model. The company cannot say which workflow the tool should improve, the information it needs is scattered across inboxes, LINE, spreadsheets and paper, and nobody owns the result. Fix those first: map the work, decide which documents and records are official, and give each process an owner. Then ChatGPT, Microsoft Copilot, Gemini or Claude has something real to work on.

Japan’s AI gap is usually discussed at national scale: fewer frontier labs, less venture capital, slower corporate uptake. Some of that is true, but it is not what decides whether a ryokan in Nagano, a bilingual service firm in Tokyo or a 30-person manufacturer gets any value from AI. What decides it is whether the business has anywhere sensible for AI to attach.

Why do AI pilots fail in Japanese companies?

Buying the software is the easy part. A company signs up for a ChatGPT business plan, Microsoft 365 Copilot or Gemini in Google Workspace. Training is scheduled, a pilot starts, a few motivated people use it and most people quietly do not. Six months later a progress report goes to management and the work looks the same.

Nobody calls it a failure, because nothing broke. The tool simply became one more object inside the existing mess.

The pattern is predictable once you look at what the tool was given to work with:

If the inquiry flow is unclear, AI drafts replies into a process nobody follows.
If the CRM is stale, AI summarizes stale data with confidence.
If the shared drive has three versions of every price list, AI search finds all three.
If nobody owns the website, AI-generated content sits in a publishing queue.
If the team does not know which workflow is official, AI speeds up the unofficial one.

None of these are model problems. A better model produces the same result faster.

What does a business need before AI can help?

The missing piece is legibility: the ability to answer basic questions about your own operation without calling the one person who remembers everything. Try these:

What systems do we use, and what is each one for?
Who holds the admin account for each of them?
Where does customer information first arrive?
Which documents are the official versions?
Where are decisions recorded?
Which vendors control infrastructure we should own, such as the domain or the website?
What breaks if one particular person leaves?

If those questions are hard to answer, adding AI gives you a faster way to produce plausible text on top of confusion. Language models are genuinely useful for drafting emails, summarizing meetings, translating between Japanese and English, classifying inquiries and turning messy notes into checklists. I use them every day. Every one of those uses depends on knowing what the workflow is, where the source material lives and who decides what happens next.

Why is this harder in Japan?

Japanese businesses tend to have a particular version of the problem.

Informal continuity. If the process works because Tanaka-san knows who to call, it is treated as a working system. It works until Tanaka-san is on leave, retires or is simply busy.

Vendor dependency that looks like stability. The website company handles the website. The tax accountant (税理士) handles the accounting software. An old IT vendor or SIer holds the domain. Each piece has a caretaker, and nobody inside the business owns the map. I wrote about this in more detail in why Japan digital transformation fails before software.

Paper and fax at the edges. Orders still arrive by fax from some customers, invoices are stamped with a hanko, and the electronic bookkeeping law (電子帳簿保存法) and the invoice system (インボイス制度) have pushed some records into digital form while leaving others on paper. AI cannot read a filing cabinet. The fax migration guide covers how to close that gap.

Two languages at once. Foreign-owned and bilingual businesses run in English and Japanese across overseas headquarters, Japanese accountants, government forms and LINE or Chatwork handoffs. Translation is one of the most useful things AI does, but only if the underlying policies and templates exist in the first place.

What should you audit before buying an AI tool?

Before I recommend any AI tool to a Japanese SME, I want the boring map.

Accounts and ownership

Who owns the domain and DNS? Which inboxes matter? Who can recover admin access to Google Workspace or Microsoft 365? Which subscriptions are paid for and which are dead weight? The subscription review is a good place to start.

Where the work actually happens

Where do inquiries arrive, who answers them and what happens after the first reply? Where does the team rewrite the same message every week? Where do customers ask the same question? Where does a monthly report take two hours because someone stitches together three spreadsheets by hand?

Where the data lives

Is customer history in a CRM, in email, in LINE, or in someone’s memory? Is there one current price list? Are the shared drives organized enough that a search returns the right file?

The answers point to the first sensible uses of AI. Usually they are small: drafting replies from an approved FAQ, summarizing CRM history before a sales call, producing bilingual versions of standard documents, or preparing a first draft of the monthly report from data that already exists.

What does a good first AI setup look like?

For most small companies it is less exotic than the sales pitch. It starts with Google Workspace or Microsoft 365 set up properly, with company-owned accounts, sensible sharing and a clean drive structure. It includes a CRM the team actually uses, a clear inquiry flow, service pages and an FAQ that answer real customer questions, and a task tool with named owners. Add a few documented automations in Zapier or Make that someone understands.

Then add AI, starting with the assistant that comes with your existing suite. Copilot sits inside Microsoft 365 and Gemini inside Google Workspace, so they work on files and email you already manage. A standalone assistant such as ChatGPT or Claude makes sense for drafting, research and analysis. On any of them, use a business plan rather than personal accounts, so conversations stay under company control and are not used for model training by default.

With that base in place, AI tools start doing useful work. A support assistant answers from a real FAQ. A sales assistant summarizes real CRM history. A bilingual assistant helps staff draft messages without inventing policy. For a practical guide to getting reliable output once you are there, see how to get reliable answers from AI tools at work.

How should a small team start?

Pick one workflow. Choose something frequent, text-heavy and low-risk, such as first replies to inquiries or meeting summaries. Avoid anything where an error goes straight to a customer without review.

Write the current process down. One page. Who does what, with which tools, using which source documents.

Fix the source material. Update the FAQ, choose the official price list, clean the templates. This step often delivers value before any AI is involved.

Run it with a human check. Staff review every AI draft for a few weeks. Note where it goes wrong and whether the cause is the tool or the source material.

Assign an owner. Someone keeps the prompts, templates and source documents current. Without that, quality drifts within months. I show what an audit of an unattended automation turns up in what 851 unattended bot runs taught me about automation.

Decide what not to automate. Some work should stay with people. I cover where to draw that line in which work should stay with people when you adopt AI.

Is Japan’s late start an advantage?

There is an optimistic version of the Japan AI story. The country arrives late, learns from other markets’ mistakes, and applies AI to real pressures such as the labor shortage. I would like that to be true. It only works if organizations can implement. If AI becomes another vendor-led pilot, another subsidized platform or another training session nobody applies, a late start is just a late start. The same applies to large regional investment, as I argue in what rural Japan needs from big technology investment.

For small businesses the practical version is simple. Before buying the next platform, map the system, own the accounts, decide what is official, remove abandoned tools and document the workflow. Then find the places where AI removes real friction.

If your business wants to use AI but the tools and data underneath are messy, a Diagnostics review will show what you actually run on, what is fragile and where AI can help. If the plan is already clear, I can build and implement it with your team. You can also start with the free Technology Risk Self-Check.

Further reading: how to evaluate an AI vendor before you buy · what 851 unattended bot runs taught me about automation · Japan’s copyright rules for businesses using AI · what belongs in a Japan SME technology stack