When a business adopts AI, the work that should stay with people is the work where someone has to be accountable: final decisions, anything said or promised to a customer, checking facts and figures, handling exceptions, and teaching junior staff how the business works. AI tools such as ChatGPT, Claude, Gemini and Copilot are good at producing first drafts, summaries, translations and sorted lists. The sensible division is to let AI handle volume and preparation while people handle judgment and responsibility, and to design that split deliberately rather than letting it happen by accident.

Much of the public conversation about AI and work is about artificial general intelligence: systems that could match people at almost any task, and the race between large technology companies to build them. That debate is worth following, but it does not tell a 20-person company in Japan what to do on Monday. What today’s tools can reliably do is narrower, and that is where decisions should be made.

Why does it matter where you draw the line?

It is tempting to think of AI as a cheap new workforce: tireless, fast and increasingly capable at cognitive tasks. Some of that is accurate. AI can now do a large share of the routine writing, research and data handling that used to fill junior roles.

The risk is treating that as a simple substitution. A business that hands all its routine thinking to software can lose three things it needs later:

  • Judgment. Someone has to notice when an answer is plausible but wrong, and that skill comes from having done the work.
  • Accountability. An AI cannot apologize to a customer, sign a contract or explain a decision to a regulator.
  • The pipeline of future managers. Senior people learned the business by doing its routine work first.

In Japan these risks arrive alongside a real labor shortage. Many small businesses are adopting automation because they cannot hire, not because they want to cut staff. That makes AI genuinely valuable here, and it makes the question of which work people keep more urgent, because there are fewer people to keep it. Care work shows the problem most clearly, as I discuss in Japan’s care technology should support caregivers.

Which tasks can AI handle, and which need a person?

A practical way to divide work is by the kind of load each task carries.

Generating and processing: good for AI

  • First drafts of emails, proposals, job descriptions and documentation
  • Summaries of meetings, long email threads and reports
  • Translation between Japanese and English for internal use
  • Sorting and classifying inquiries, receipts or survey responses
  • Turning notes into checklists, tables or structured records
  • Preparing data for a report that a person will review

Evaluating and deciding: keep with people

  • Approving anything sent to a customer, partner or government office
  • Pricing, discounts and contract terms
  • Hiring, performance and anything affecting an individual employee
  • Handling complaints and relationships that are going badly
  • Deciding what to do when the usual process does not fit
  • Confirming facts, figures and regulatory requirements before they are relied on

The line moves as tools improve and as your team builds trust in a particular workflow. Move it deliberately, one workflow at a time, after checking results, rather than because a new model was announced.

What happens to junior staff when AI does the routine work?

This is the question I think most businesses underestimate. Traditionally, junior people learned by doing rote work: drafting the first version of a proposal, reconciling the spreadsheet, reading the contract. In a law firm, a junior associate reviewing contracts learns how deals are structured. If AI does that review in seconds, the junior may never learn it.

Japanese companies have long relied on on-the-job training (OJT) rather than formal programs. OJT depends on there being routine work to learn from. Remove that work and the training disappears with it, often without anyone noticing until a senior person leaves.

A few ways to keep the pipeline working:

Have juniors review AI output, not just receive it. Checking a draft against the source and explaining what is wrong teaches much of what writing it would have.

Keep some work manual on purpose. A new hire should build a few reports or proposals from scratch before using AI to speed them up.

Write down what experienced people know. Procedures, decision rules and the reasons behind them are useful to staff and to AI tools alike. The same documentation serves both.

I write more about building skills around real work in why reskilling is hard in Japan and why that is an opportunity.

How does AI change fraud and trust?

One risk arrives whether or not your business adopts AI. Generating convincing fake emails, invoices, voices and video is now cheap. Phishing emails in natural Japanese, fake supplier requests to change bank details, and voice calls imitating a manager are all realistic threats. Japan’s IPA keeps a page on business email compromise with case examples.

People and process are the defense here, not more technology alone:

  • Confirm any change to payment details through a second channel, such as a phone number you already have on file.
  • Require two people to approve payments above a set amount.
  • Use multi-factor authentication on email, banking and core business accounts.
  • Tell staff that an urgent request from a senior person is exactly what an attacker would send.

Will customers start sending AI agents to buy?

Personal AI assistants that search, compare and book on someone’s behalf are already appearing, including apps inside ChatGPT. If that grows, some of your customers will be represented by software that reads your website and compares your prices before a person ever sees you.

For small businesses the preparation is ordinary good practice: clear service pages, accurate prices and availability, an FAQ that answers real questions, and booking or inquiry forms that work. Businesses whose information is only in PDFs, Instagram posts or someone’s head will be hard for people and software alike to find. I cover the connection side in what to check before connecting business apps to ChatGPT.

Is narrow AI the right approach for most businesses?

For almost every small and mid-sized business, yes. The useful approach is not to wait for a general intelligence that can run the company, but to apply specific tools to specific jobs: summarizing, drafting, classifying, translating. This fits the Japanese habit of kaizen well. Improve one process, measure it, keep what works, then move to the next.

That approach also keeps the business in control. You know which tasks AI handles, who checks the result, and what to do if a tool changes or disappears.

A simple plan

List your recurring work. Mark each task as generating, processing, evaluating or deciding.

Automate preparation first. Start with tasks where AI prepares and a person approves.

Protect training. Decide how new staff will learn the work that AI now does.

Tighten verification. Put payment and identity checks in place before you need them.

Review every few months. Move the line when the evidence supports it.

If you are working out how AI fits into your team and processes, I can help map the work and set up the tools with you through Ongoing Stewardship. For the foundations that make this possible, see why AI adoption stalls in Japanese SMEs.

Further reading: how to evaluate an AI vendor before you buy · how to get reliable work answers from AI tools · Japan’s care technology should support caregivers · TetsuClaw: AI work OS for operators in Japan