Demographics do the work. AI is secondary.
Employment holds roughly flat to around 2031, then turns down — about 6% lower by 2040, at roughly 60.6 million across the industries modelled. Even under the model's most disruptive AI-substitution scenario, about 80% of that decline is demographic. AI accounts for the remaining 20%.
The flows are blunter than the stock. Retirements run at roughly 1.3 million a year against fewer than 600,000 new entrants. Net annual shrinkage of the labour force goes from about 172,000 in 2026 to about 757,000 by 2040.
There is one odd dip in that curve, in 2031, and it is not a data error. The hinoeuma (丙午) superstition depressed births in 1966 by around 25%. That small cohort turns 65 in 2031.
The three buffers that carried the last decade are all decelerating. Since 2013, women's employment is up about 16% and employment of over-65s about 48%; foreign workers have multiplied from a low base and are projected to reach roughly 8.5% of employment by 2040. Each of those is real. Each is approaching a natural ceiling.
AI relieves the sectors that were not your bottleneck
AI's effect concentrates in information and communications, finance and insurance, professional and technical services, and education. Read that carefully: those sectors do not go into surplus. Japan is short of people almost everywhere. These are simply the sectors most exposed to AI relative to how short they already are, and the effect shows up as softer hiring pressure, not redundancies.
Where it does not show up is healthcare and social care (医療・福祉), construction, and transport — the three tightest sectors in the country. A back-office hour freed in Marunouchi does not become a care worker in Saitama or a site supervisor in Nagoya.
Semiconductors and electronic components run the other way. Sitting upstream of AI, employment there grows by around 20%.
Unemployment stays low. That is the problem, not the reassurance.
Roughly 2.5% in 2025, about 3.0% in 2040 on the AI baseline, and 4.8% even under full substitution. Information and communications tops out near 10% in the harshest scenario, because the structural engineer shortage absorbs most of the displacement.
So the tightness never presents as visible unemployment. It presents as requisitions that stay open in care, construction and the skilled trades, alongside white-collar staff who stay put because there is nowhere adjacent for them to go.
Japanese workers do move. Just not toward the shortage.
This should reset how you think about reskilling (リスキリング).
About 60% of Japanese job changers move across occupations, on Indeed's resume data. Mobility is not the constraint. Direction is.
Nursing is the clean illustration. 85% of moves into nursing come from nursing, and 81% of those leaving a nursing role move to another one. The rest go to caregiving, retail and clerical work — the low-barrier destinations. Inflow from finance or IT is negligible. Licensing walls, membership-based employment (メンバーシップ型雇用) and seniority pay (年功賃金) hold those flows in place.
Measured against an idealised frictionless market, reallocation friction alone accounts for about 1.1 percentage points of unemployment in 2040 under the substitution scenario — roughly 680,000 people.
The retirement-extension trap
This is the finding I would put in front of a board.
The model tests two policy counterfactuals. Loosening dismissal rules (解雇規制緩和) raises total turnover by about 30% but lifts net inter-industry reallocation only 3–7%, and nudges unemployment up rather than down. In a shortage economy the shortage is already pulling people toward growth sectors. Easier dismissal mostly produces more churn. The binding constraints are matching and reskilling.
Extending working life (定年延長) is the genuinely powerful lever: about 2.66 million more people in employment by 2040. It also carries a sting. By filling the shortage back in, it thins the very buffer that has been quietly absorbing AI displacement. Under substitution, 2040 unemployment moves from 4.8% to 5.7%, AI-attributed unemployment from roughly 1.1 million to 1.6 million, and AI's share of the total employment decline jumps from about 20% to nearly 60%.
The amplification runs both ways — complementary AI offsets around five times as much decline under the same settings. Extending senior participation is sound policy. It also raises the stakes on whether AI arrives as a complement or a substitute.
And Japan has barely started
18% of workers here use AI at work, and AI mentions in job postings sit among the lowest of the major economies. The survey behind those numbers points at employers, not workers, as the brake.
In a shortage economy the usual reason to hesitate — job losses — largely does not apply. Most of what AI automates gets absorbed by the vacancy you could not fill anyway.
Five moves for hiring leaders
- Plan headcount by sector, not off the national average. The aggregate 6% decline tells you almost nothing about your own seat. Which of the 16 industries you sit in decides whether AI softens your pipeline or leaves it untouched.
- Treat AI as a hiring-pressure tool, not a headcount-reduction tool. In white-collar functions the honest business case is fewer open requisitions and less recruiting spend, not severance. Build the case that way and adoption stops being a threat internally.
- If your bottleneck is physical or human-contact work, buy retention (定着), not pipeline. There is no inflow coming from the AI-exposed occupations. Everything you keep is worth more than anything you can recruit.
- Build adjacent-move ladders inside your own organisation. Cross-occupational moves already happen at scale in Japan; they just default to low-barrier destinations. Staged qualifications and internal transparency are what redirect them.
- Push AI adoption from the top. The brake is employer sponsorship, not worker resistance. If your workforce is going to shrink either way, the productivity gain is the only variable you actually control.
The next fifteen years in this market will not be decided by a wave of redundancies. They will be decided by whether people can move to where they are needed — and on current evidence, they largely cannot.