What ChatGPT can and cannot do for your Japan project

If you are planning anything in Japan, you have almost certainly already asked an AI about it. That is a sensible place to start, and for a surprising amount of the work it is genuinely enough. This article is about the part where it stops being enough — written by a Japanese company that does the remaining part for a living, so read it with that in mind.

We are not going to tell you AI is useless for Japan work. It isn’t. We use it ourselves. What we want to give you is a clear line: here is what a language model can do from your desk, and here is what requires a person standing in a government office in Tokyo.

What can AI genuinely do for a Japan project?

More than most people expect. If your question is about the shape of the Japanese market rather than a specific fact about a specific company, a good model will get you a long way.

It can explain how Japanese distribution layers work, why a foreign brand usually ends up with an importer rather than selling direct, and roughly what your category looks like. It can draft the first version of a partner brief. It can translate a Japanese company profile well enough for you to understand it. It can summarise a regulation in plain English so you know which questions to ask a lawyer.

For orientation, AI is excellent and nearly free. Use it. Skipping that step and paying a consultancy ¥1,500,000 to tell you what a model would have told you in an afternoon is a bad trade.

Where does AI stop working?

It stops at four boundaries: documents that are not on the web, actions that require a physical presence, language that has to convince a human, and liability. Everything below is a version of one of those four.

Why can’t AI just look up a Japanese company?

Because the useful part of the Japanese corporate record is not on the open web, and no amount of model capability changes that.

Japan does have a free public registry. The National Tax Agency’s corporate number site publishes what it calls the basic three items: the company’s name, the address of its head office, and its 13-digit corporate number. That is the complete list. It is genuinely useful — it proves a legal entity with that name is registered at that address — and any AI with web access can find it.

What it does not contain: capital, directors, date of incorporation, stated business purpose, share transfers, or history of name and address changes. Those live in the commercial register at the Legal Affairs Bureau, and you have to request them, per company, and pay for them. As of 1 April 2025 a certificate of registered matters costs ¥600 at the counter, ¥520 requested online and posted, or ¥490 requested online and collected in person. An uncertified copy of the same data through the online Registration Information Provision Service costs ¥330 as of 1 April 2026 — cheaper, but with no seal, so a bank or a court will not accept it.

None of that is behind a paywall in the way a subscription database is. It is behind a process. Someone has to make the request. That is why “just look it up” is not an instruction you can give a model.

Why doesn’t AI find Japanese financial statements?

Because for most Japanese companies, they were never published anywhere a crawler goes.

Listed companies file properly and their filings are online. Below that, Japan’s Companies Act obliges every joint-stock company to publish an annual financial notice after its shareholders’ meeting — but the permitted channels include the Official Gazette (Kanpo) and daily newspapers, not just a website. A notice printed in the Gazette is technically public and practically invisible: it is not indexed the way a Companies House filing in the UK is.

So when an AI tells you it cannot find a mid-sized Japanese manufacturer’s revenue, it is not failing. It is accurately reporting that the number is not on the web. Getting it means commercial credit databases, or asking the company, or inference from other sources — all of which are research tasks, not lookups.

Why does AI-written Japanese get ignored?

Because Japanese business writing encodes the relationship between sender and recipient, and a translation engine has no view on what that relationship is.

The clearest example is the pair 御社 and 貴社. Both mean “your company”. 御社 is what you say out loud; 貴社 is what you write. Translation tools pick one, often the wrong one for the medium, and a Japanese reader registers it instantly — not as an error exactly, but as a signal that whoever wrote this does not do business in Japanese.

Multiply that by the opening formula, the order of self-introduction, the level of honorific used for a first contact versus a continuing relationship, and where the actual request sits in the message, and you get an email that is grammatically correct and reads as machine-generated. Machine-generated cold email in Japanese has one destination, and it is not a reply.

This is not an argument against using AI to draft. Draft away. The argument is that the last pass has to be done by someone who is deciding, consciously, how formal to be with this particular recipient — because that decision is the message.

Why can’t AI make the phone call?

Because a large share of Japanese B2B first contact still happens by telephone, and no one has yet built a bridge from an AI agent to a Japanese company’s main switchboard that a receptionist will cooperate with.

Email to a general info@ address at a Japanese SME is frequently unread. The route that works is calling the main number, explaining who you are, and asking to be put through to the right department. That is a conversation with a human gatekeeper, conducted in Japanese, with a real company name attached to it. It is completely ordinary and completely outside what a model can do.

What about the things that need a body in a room?

Trade shows, retail shelves, factory floors, and the registered address of a company that may or may not be operating from it.

Japanese B2B still runs a serious trade-show circuit, and for many categories that is where relationships actually begin. A stand at a Tokyo exhibition is a real place with real people handing out real business cards. Someone has to be there. Similarly, if you want to know whether a supplier’s premises match what they claim, the answer is a photograph taken by someone standing outside the building.

And liability?

This is the one people forget. At some point a Japanese counterparty needs someone to be responsible.

Signing an agreement, attending a meeting as your representative, receiving samples, being the named contact on a form — these require an entity that can be held to account. An AI cannot be a party to anything. Increasingly, neither can an anonymous freelancer: for certain arrangements a Japanese counterparty will expect a registered company on the other side, with a corporate number they can look up.

So what is the honest division of labour?

Use AI for everything up to the point where the task requires a document you must request, an action in physical space, a sentence that has to persuade a Japanese reader, or a signature.

Concretely, we would suggest this split:

• Do it with AI: market orientation, understanding channel structure, reading and summarising Japanese public information you have already obtained, first drafts of anything, understanding what a regulation broadly requires.
• Do it with a person in Japan: obtaining registry documents, verifying a company actually operates, writing anything a Japanese recipient will judge you by, phoning, attending, negotiating, signing.

The useful mental test is: if this goes wrong, who is accountable? If the answer is “nobody”, you are still in the AI half. If the answer needs to be a name, you have crossed the line.

When is it worth outsourcing this?

When the cost of being wrong exceeds the cost of asking. That is usually earlier than people think.

A verification report costs a few hundred dollars. Sending samples, paying a deposit, or signing a distribution agreement with a company that is not what it appears to be costs considerably more — and the mistakes tend to surface after the money has moved. The same logic applies to outreach: twenty carefully written approaches from a Japanese company get a different response rate than two hundred translated ones, and the difference is not marginal.

We built this desk for exactly the tasks in the right-hand column of the table above — the ones that need a registered Japanese company, at a fixed price, without a six-month contract. Everything in the left-hand column, you should keep doing yourself.