AI Agents Remove Learning Tasks. Japan's Curve.

The question executives keep asking about AI agents is a headcount question: how many roles go away, and when. The usual response is a freeze on junior hiring. The cost of that shows up years later, when the seniors who should have come up through those roles do not exist.

What the data shows so far is that AI agents are taking tasks rather than whole jobs, and mostly the low-stakes, repetitive tasks that organisations have always used to teach judgment to people who did not have it yet, such as first-pass document review, initial data extraction and reconciliation, draft code, and first-cut financial models.

This is the part of the picture where Japan currently differs from the United States, for reasons that have more to do with labour law than with preparation.

The learning curve is what is being automated

PwC’s 2026 Global AI Jobs Barometer describes a labour market splitting into two paths rather than shrinking along one. In the sectors most exposed to AI, early-career postings have gone flat. The entry-level roles that remain have changed shape: they carry responsibility that used to sit a level or two up, and that category has grown substantially since 2019.

The work that used to justify hiring someone with no track record, where the real product was the trained employee rather than the output, now has a cheaper substitute. The entry-level work that remains demands judgment that entry-level people have not had a chance to build.

Japan is on a different clock

At Tech for Impact Summit 2026 the sharpest version of this argument came from an investor.

Yumiko Murakami, General Partner at MPower Partners Fund and formerly Head of the OECD Tokyo Centre, drew the contrast on the Main Stage during the Learning in the Age of AI panel:

“If you look at the U.S. right now, both startups and established companies, they are laying off huge, huge number of entry-level jobs. That’s not happening in Japan. That’s not happening at all in Japan.”

The reason is not that Japanese firms are better prepared. Japan’s labour market is rigid and demographically thin. Dismissal is legally and culturally difficult, and most large employers are trying to fill roles, not shed them. International commentary has spent three decades calling that rigidity Japan’s biggest weakness. In 2026 it is slowing the layoffs.

Murakami put Japan’s window at a handful of years, and said it only pays off if companies use it to retrain rather than to postpone. Her summary of Japan’s position in 2026 was that it is later than the U.S., not immune.

Adoption is not usage

Maiko Kojima, founder and CEO of the enterprise AI company Crafter and author of a book built on conversations with more than a thousand corporate AI leaders, described a gap between what companies have bought and what their people do with it. Citing NTT Docomo’s enterprise survey, she put Japanese corporate AI adoption at roughly 50–60%, up from something closer to 20% two years earlier.

Then she described what that adoption looks like inside the building:

“They adopted AI, but not so many people, not so many employees are using their own AI. If they use it, it’s almost like very limited use cases. They only use it for surveys. They only use it for brainstorming. Those use cases are very good. But it has to be adopted on their own workforce, like the AI agent.”

Adoption figures count purchases. A company can sit inside the 60% with single-digit active usage: a licence deployed, a pilot completed, a press release issued, and a workforce that works exactly as before.

Such a firm has paid for the tools, so it feels justified in slowing junior hiring, but it has not restructured any workflow, so nothing has been automated and no capacity has been freed. The junior roles have been cut on paper and nothing has replaced them.

When Kojima’s team surveys employees after a rollout, the request that comes back is the same across companies: training and reskilling. The training that works is not productivity training. It is critical thinking and source evaluation, teaching people to interrogate what a system hands them.

The dojo problem

Satoshi Hirose sees the same thing in a classroom.

Hirose, Dean of GLOBIS University’s Graduate School of Management, described his MBA classes. He asks a deliberately fragmented question. Thirty seconds later a student offers a polished answer. He asks how they arrived at it, and most of the time the student says plainly that it is AI’s answer.

“Gradually, they are students who are, in a lack of a better word, like, prisoned by, hostage by AI.”

Hirose was careful to say this was not laziness. The output was fine. What never happened was the reasoning that was supposed to be built while producing it.

Companies have their own version of this. Entry-level work was where judgment got built under low stakes and supervision, and if agents do that work, the work still gets done while the trainee goes untrained.

What to restructure

If the tasks that used to train people are now done by agents, judgment has to be built somewhere else, and that is a design problem rather than a headcount one.

One practical step is to go through each task an agent now handles and ask what it used to teach, and to whom. Some of those tasks taught nothing and could have been automated years ago. Others were the whole training pipeline, and those need a replacement before the junior role disappears.

If juniors are going to supervise agent output rather than produce first drafts, then critical evaluation becomes the job, and it has to be taught explicitly. Kojima’s survey data says employees are asking for exactly that. Most corporate AI training does not currently cover it.

The metric to watch is active usage per employee, by team and by use case, rather than the adoption figure. It is the only number that shows whether a workflow has changed and whether any capacity has been freed.

Most companies are already making this decision, usually by default, in the form of a junior req that does not get reopened. Japanese companies have a few more years than most to make it on purpose.

The perspectives above are drawn from the Learning in the Age of AI panel at Tech for Impact Summit 2026, moderated by Harry Dempsey of the Financial Times. The full session recap is available in our insights archive.

Tech for Impact Summit returns to Tokyo on May 18–19, 2027, as a partner event of SusHi Tech Tokyo. The summit is invitation-only, convening leaders across business, policy, and culture around the technologies that decide the next decade. If the questions in this briefing are ones your organisation is working through, we would welcome your interest. Request an invitation.