ENGLISH Contact
BUSINESS── ARTICLE 038── ── 15 MIN READ

Did AI really halve Japan's office jobs? We counted 14 years of government job data

Headlines in Japan say job ads for office clerks have been cut in half. DATA TIMES stitched together the government's occupation-level job data from Hello Work, Japan's public employment service, from 2012 to September 2026. New openings for general office clerks are down 10.3% since the year before ChatGPT, but all openings are down 8.4%. Strip out the economy and a gap of 4.9% below the expected level opened only in 2026. We set that against the private ad counts behind the 'halved' story, record clerical employment, and a Dallas Fed study, to show how much of 'AI is taking office jobs' the data can actually support.

In a dark office archive, a towering stack of white job-posting sheets is split down the middle by a beam of golden light, the right half collapsing into the air. On the left, DATA TIMES editor-in-chief Kaito Ogasawara holds a magnifying glass and looks at the camera with suspicion. On the right, the words OFFICE JOBS “HALVED”? and, in huge gold serif letters, THE TRUTH.
Job ads for full-time office clerks in Japan fell by half in a single year. Since the autumn of 2025, that number has kept turning up in the news.

It comes from the monthly count published by the Association of Job Information of Japan, an industry body of job-site and job-magazine companies. In January 2026, ads for full-time "office work" were down 53.6% from a year earlier. When the slide first showed up in September 2025, Reuters said the main cause appeared to be AI automating office tasks. The timing fits the moment generative AI started drafting documents and filling in spreadsheets, so the number is easy to read as "AI has started taking office jobs."

Our editor-in-chief, Kaito Ogasawara, had one question when he saw it: does the government's own data say the same thing? Let's count.

About one in five people working in Japan has an office job. If openings for that work had really halved, it would matter to every student looking for a first job, every worker thinking about a move, and every company planning its hiring.

So we took the government's monthly job data by occupation from Hello Work, Japan's public employment service, and joined it into one series running from March 2012 to September 2026. New openings for general office clerks are down 10.3% from the year before ChatGPT launched. But openings across all jobs are down 8.4% too. And once you strip out the ups and downs of the economy, a gap of 4.9% below the expected level only began to open in 2026.

Part of a reference table from Japan's Ministry of Health, Labour and Welfare for September 2026, listing new openings, active openings, new applicants, active applicants, referrals, hires and job-to-applicant ratios by occupation. The ratio is 1.08 for all occupations, 0.37 for clerical workers and 0.30 for general clerks, against 2.55 for social welfare specialists.
Hello Work's occupation table for September 2026 (regular jobs, including part-time). The active job-to-applicant ratio for general clerks is 0.30: three openings for every ten people looking. This article counts the long-run data behind tables like this one. (MHLW reference tables, September 2026)
WHAT IS

Hello WorkPublic Employment Security Office

Japan's free, nationwide public job center for job seekers and employers

  1. 1947A public job officeOn April 17, 1947, Japan's employment offices were renamed Public Employment Security Offices.
  2. 1990Named Hello WorkThe offices adopted the nickname Hello Work, chosen in a public contest to sound friendlier.
  3. 2021.09Online sign-upJob seekers could register and apply online without visiting an office.
  4. NOWSeptember data outOn October 30, 2026, it published a 1.19 job-to-applicant ratio and openings by occupation.

Hello Work is the everyday name for Japan's Public Employment Security Offices. The government runs 544 of them across the country, matching people who need work with companies that need people, free of charge. The job-to-applicant ratio you see in Japanese economic news is calculated from the openings and job seekers registered here.

It is big. In September 2026 alone, employers posted about 710,000 new regular openings, part-time included. The labor ministry publishes the data every month as its Employment Security Statistics, broken down by occupation, and anyone can download it.

It does not see everything, though. According to the Cabinet Office's 2026 economic white paper, the share of new hires who found their job through Hello Work fell from 24% in 2013 to 16% in 2024. We keep that in mind throughout.

ChatGPT launched on November 30, 2022, almost four years ago. Working with AI and data every day, we had been wondering the same thing as we drafted documents and tables with AI: what is it doing to office work?

We checked three things. Did openings for office jobs really fall? If so, did they fall by more than the economy explains? And can we pin that on AI?

01

Office openings fell 10%, but all openings fell 8%

Chapter opener: a long row of public job-center notice boards at dusk, with job cards missing from every board alike and one board in front lit in gold. Text reads OFFICE JOBS, DOWN 10%. WHY?

First, the data and how we counted it.

The labor ministry publishes Hello Work's openings and job seekers by occupation on e-Stat, Japan's government statistics portal, as long-term Table 21. Because the occupation categories changed along the way, the data comes in two files: one using the ministry's own classification from March 2012 to March 2023, and one using Japan's Standard Occupational Classification from April 2022.

The two files overlap for the year from April 2022 to March 2023. We compared new openings for general clerks, accountants and bookkeepers, construction workers, drivers and all jobs combined, and every month matched to the last opening. So we used the older file up to March 2022 and the newer one from April 2022, giving one continuous record.

We counted new openings: the jobs employers newly posted with Hello Work in a given month. It is the closest number to how many people companies want to hire. We used regular jobs including part-time ones, and summed 12 months at a time to remove the seasonal swing (openings jump around the start of Japan's fiscal year in April, for example).

The comparison is between the year just before ChatGPT (December 2021 to November 2022) and the latest year (October 2025 to September 2026).

Horizontal bar chart of the change in new job openings by occupation: general clerks −10.3%, all jobs −8.4%, construction −9.0%, care workers −6.0%, drivers −4.4%.
Hello Work new openings (regular jobs including part-time), 12-month totals, comparing the year before ChatGPT (December 2021 to November 2022) with the latest year (October 2025 to September 2026). Chart by DATA TIMES. (MHLW job statistics, compiled by DATA TIMES)

New openings for general clerks fell from 648,391 to 581,721, a drop of 10.3%. That is roughly 54,000 a month down to about 48,500.

Over the same period, though, Hello Work's new openings across all jobs fell from 9.26 million to 8.48 million, or 8.4%. Hands-on jobs that AI can hardly touch fell too: construction and mining by 9.0%, care work by 6.0%, and truck, taxi and other driving jobs by 4.4%.

OccupationYear before ChatGPTLatest yearChange
General clerks648,391581,721−10.3%
Accounting clerks81,08876,192−6.0%
All jobs9,256,0238,477,183−8.4%
Construction and mining499,686454,813−9.0%
Care workers870,394817,919−6.0%
Drivers425,964407,338−4.4%
Cooks528,795457,426−13.5%
Sales (all)865,071738,175−14.7%

Source: compiled by DATA TIMES from MHLW Employment Security Statistics, long-term Table 21. New openings for regular jobs including part-time, 12-month totals. Year before ChatGPT: December 2021 to November 2022. Latest year: October 2025 to September 2026.

General clerks fell a bit more than the middle of the pack. Cooks (down 13.5%) and sales jobs (down 14.7%) fell much further.

AI alone cannot explain why openings are falling everywhere. The Cabinet Office white paper points to one reason: fewer job seekers and employers are going through Hello Work at all.

Take out the overall decline and general clerks fell just 2.0% more than everything else. Office openings really are down 10%, but most of that is the whole market shrinking.

02

Take out the economy, and a 5% gap opened in 2026

Chapter opener: a large brass balance in a dark room, a stack of white job cards on one pan and a glowing gold weight on the other, the beam slightly tilted with a thin line of golden light. Text reads MINUS THE ECONOMY, A 5% GAP.

Is "only 2% worse than everything else" the end of the story? Next we tracked general clerks' share of all new openings over time.

Line chart of general clerks' share of Hello Work's new openings. From 2014 to 2019 it stays within a 7.1–7.3% band, drops sharply during COVID in 2020–2022, recovers to 7.4% in 2023 after ChatGPT's launch, then falls steadily to 6.9% in the latest year.
General clerks' share of new openings (regular jobs including part-time), calculated on 12-month totals. The grey band shows the 2014–2019 range of 7.1–7.3%. Chart by DATA TIMES. (MHLW job statistics, compiled by DATA TIMES)

For the six years before COVID, from 2014 to 2019, the share barely moved outside 7.1–7.3%. Boom or slowdown, about one in every 14 openings at Hello Work was for a general clerk. It was a remarkably steady job.

COVID hit office hiring first, and the share sank to 6.69% in November 2021. Then, after ChatGPT appeared, it went up, not down, reaching 7.39% in August 2023, its highest since early 2013. Right after generative AI arrived, office openings were actually recovering.

The change came later. The share has fallen almost without pause since the summer of 2023, to 6.86% in the latest year. Leaving aside COVID in 2021–2022, that is the lowest since the record begins in 2013.

Year-on-year numbers point the same way. In 10 of the 12 months from October 2025 to September 2026, openings for general clerks fell faster than openings overall. In September 2026, all openings were down 0.9% while general clerks were down 7.4%.

A share is easy to read, but it can be distorted when total openings swing a lot. So we went one step further and calculated an expected level based on the economy.

Here is the method. Using the 72 months from January 2014 to December 2019, before COVID, we ran a log-linear regression of general-clerk openings on total openings plus month-of-year effects. When total openings rise 1%, general-clerk openings rise about 1.03%, and the model misses individual months by roughly ±2%. Feed in total openings from 2023 onward, and you get the number of general-clerk openings you would expect given the economy and the season.

Bar chart of how far new openings for general clerks were above or below the level predicted by the economy each year: +3.2% in 2023, +1.4% in 2024, −1.1% in 2025 and −4.9% in January to September 2026, about 2,500 fewer openings a month than expected.
Actual new openings for general clerks compared with the level predicted from total openings and the season, using the 2014–2019 relationship. 2026 covers January to September. Chart by DATA TIMES. (MHLW job statistics, DATA TIMES estimate)

The pattern is clear. In 2023, openings were 3.2% above the expected level, and in 2024 still 1.4% above. In 2025 they slipped 1.1% below, and from January to September 2026 they were 4.9% below. That is about 2,500 general-clerk openings a month that the economy says should exist but don't.

We ran the same calculation for other occupations.

Occupation202320242025Jan–Sep 2026
General clerks+3.2%+1.4%−1.1%−4.9%
Accounting clerks+3.5%+5.7%+0.3%−5.0%
Care workers+6.8%+9.6%+13.0%+17.4%
Construction and mining+19.3%+18.5%+22.0%+26.1%
Drivers+5.2%+6.1%+7.3%+8.3%

Source: DATA TIMES estimate from MHLW Employment Security Statistics, long-term Table 21. For each occupation, a log regression of its monthly new openings (regular jobs including part-time) on total openings and month effects, fitted on 2014–2019; each year's actual total is compared with the predicted total.

Both kinds of office work, general and accounting, came in about 5% below expectations in 2026. Care, construction and driving, by contrast, keep running far above them. That is a different story from AI: over the latest year there were 3.8 openings per applicant in care work and 5.0 in construction. Employers cannot find people, so openings pile up.

So far we can say two things. Openings for general clerks have started to fall short of what the economy explains, and the gap widened in 2026. But it is still about 5%, nowhere near "halved."

03

Why 'halved' and 'down 10%' disagree

Chapter opener: on the left, a pile of glossy job flyers torn exactly in half; on the right, a ledger and a ruler under a desk lamp, with a line of golden light between them. Text reads HALF VS. 10%, WHY THEY DIFFER.

So what was the "halved" number? We laid out the industry association's published spreadsheets month by month.

According to the association's own notes, the figures come from weekly ad data supplied by 15 major job platforms run by volunteer member companies, re-sorted by occupation and averaged by week. Temp-agency and recruitment-agency listings are excluded. The association itself warns that ad counts "do not represent the number of job openings themselves."

Laid out in a row, the office-work numbers show a sharp step. Full-time office ads fell from about 340,000 in March 2025 to about 230,000 in July and about 170,000 in September. Counting all employment types, total ads fell 4.8% from August to September 2025, while office ads alone fell 18.9% in that single month. Since then, full-time office ads have sat flat at roughly 150,000–170,000 through September 2026.

And the year before went the other way. Office ads (all employment types) in January to March 2025 were 26–44% higher than a year earlier. A sudden drop after a sudden jump makes any year-on-year comparison look like a collapse.

IT ads up fourfold in a year, all ads up a third. No other statistic suggests Japan's job market grew anything like that. Hello Work's new openings fell over the same year, and the national job-to-applicant ratio stayed flat at around 1.2.

Jumps like these tend to appear when the set of participating platforms changes, or when the rules for which occupation an ad counts under shift. The association's materials do not explain the September 2025 step. Either way, this "halved" cannot simply be read as "half of office jobs are gone."

In the government data, the drop in openings for general clerks is about 5% once the economy is stripped out. The "halved" figure from private ads is ten times bigger.

There is one more angle: the number of people actually working in office jobs, from the Statistics Bureau's Labour Force Survey.

In the Labour Force Survey's occupation table, the number of clerical workers rose by 470,000 between 2022 and 2025, to 14.48 million, the highest since this classification began in 2009. The number of people doing office work has not fallen.

Openings are the seats being added for newcomers. Employment is the number of people already seated. More and more people are sitting in office seats, but fewer new seats are being added than the economy would suggest. That is the shape of what is happening now.

Two caveats belong here.

Hello Work is not the whole job market. If employers moved some general-clerk openings from Hello Work to private job sites or recruitment agencies, that shift is mixed into the 4.9% shortfall. Our data cannot separate the two.

And since September 2021, Hello Work also counts job seekers who registered online without visiting an office. According to the labor ministry, that affects job-seeker counts, hires and ratios, but not the new openings this article mainly uses.

04

In the US, a gap of the same size showed up two years earlier

Chapter opener: a night view from high above the Pacific, with the lights of an American city on the right and Tokyo on the left, joined by a golden arc of light. Text reads AMERICA SAW IT, 2 YEARS EARLIER.

What have researchers abroad found? The closest recent approach comes from economists at the Federal Reserve Bank of Dallas, published on September 1, 2026.

The Dallas Fed analysis counted online job ads in Texas using data from Lightcast, a labor-market data firm. It scored each occupation by the share of its tasks that generative AI can automate, based on task data published by the AI company Anthropic, then compared ads for more-exposed and less-exposed occupations within the same industry.

The findings:

  • Ads for more AI-exposed occupations fell about 5% relative to less-exposed ones by the end of 2023, and about 8% by the first quarter of 2025.
  • Among established firms, those more exposed to AI cut their postings by 8–9% by early 2026.
  • Across all Texas job ads, AI exposure reduced postings by about 1.8% in 2024 and 2.6% in 2025.

Set next to Japan's Hello Work numbers, the sizes look strikingly similar. In Texas, a gap of about 5% had appeared by the end of 2023, a year after ChatGPT. For Japan's general clerks, a gap of the same size is appearing in 2026.

The two are measured differently, though. Texas uses online job ads; Japan uses public job-center openings. Texas compares exposed jobs with less-exposed jobs; our Japanese calculation compares clerks with the economy as a whole. The sizes are fair to put side by side as a rough guide, not as the same measurement.

It is worth adding Japan's official view. On September 29, 2026, the labor ministry's White Paper on the Labour Economy took AI and the labor market as its theme. Hello Work's job-to-applicant ratio for clerical work averaged 0.46 from 2016 to 2019 and 0.44 from 2022 to 2025. The white paper concluded that there is no evidence so far that AI is causing large job losses.

That fits our numbers. Through 2025, the yearly gaps were small (+3.2%, +1.4%, −1.1%) and roughly cancel out. The gap opened clearly only in 2026, after the white paper's period ends.

Views also differ in the US. In May 2026, researchers at the Federal Reserve Bank of New York looked at the same Lightcast ads and found that the relative slowdown in postings for AI-exposed occupations began before ChatGPT, concluding that AI is not the main driver of slower hiring. Neither the Texas 5% nor the Japanese 5% can be put entirely down to AI.

Still, there are good reasons the gap took longer to show up in Japan.

The Cabinet Office's economic white paper reports a survey of where companies use AI. For routine document drafting, about 70% of companies are using or considering AI to replace the work, counting future plans. Scheduling and accounting, finance and tax work are also areas where companies already tend to hand part of the work to AI. For physical work such as transport and delivery, the figure is only 30–40%.

The same white paper finds that demand for office work holds up even in downturns and that employers feel less short of office staff than of other workers. Jobs that no one is rushing to fill tend to change first through "don't replace people who leave" when a new tool arrives. A gap that opens slowly and steadily, as Japan's has, is what you would expect.

One more number stands out. Between 2022 and 2025, wages for office work rose faster than for any other occupation group, according to the Cabinet Office. If AI were making office workers redundant, you would expect their pay to lag. The white paper says AI substitution is not pushing office wages down. The labor ministry's white paper, which looks in finer detail, does show pay for "general office clerks" rising slightly less than the overall average, so differences may be emerging within office work depending on what the job involves.

Putting it together:

  • What fell
    New openings for general clerks are 10.3% below the year before ChatGPT.
  • Where there is a real gap
    Even after stripping out the economy and the season, openings in January to September 2026 were 4.9% below the expected level, and the gap has widened every year since 2025.
  • How far we can blame AI
    The timing and size match what US research attributes to AI, and about 70% of Japanese companies are considering AI for routine document work. But Hello Work data alone cannot separate AI from other causes, such as openings moving to private platforms. The government white paper covering data through 2025 also found no sign of large AI-driven job losses.
EDITOR'S TAKE ── How editor-in-chief Ogasawara reads it

Office jobs are not vanishing. The way in is getting narrower

The government data does not back up 'halved.' But lay 14 years side by side, and a small, clear change has begun to show in office openings.

VIEW01

It is too early to say AI has taken office jobs. 'Not being replaced' has already begun

So, is the headline true? Office job openings have not halved. But something real has started.

In the government data, openings for general clerks are about 5% lower than the economy explains. The number of people in office jobs is at a record and their pay is rising. Calling that "AI taking jobs" goes too far.

At the same time, a share that sat between 7.1% and 7.3% for six straight years has been falling for three. Rather than cutting staff, companies have started leaving seats empty when people leave, and running the work with AI and better systems. That is how we read the numbers today.

VIEW02

What narrows is the way in for newcomers

The job-to-applicant ratio for general clerks is 0.30: three openings for every ten people looking. Long before AI, office work was one of the hardest doors to get through at Hello Work. The Cabinet Office notes that office jobs, which need few special qualifications and hold steady through downturns, have always attracted many applicants.

What AI changes is how wide that door is. People in office jobs today are not suddenly losing their seats; it is the seats for newcomers that are not growing with the economy. In the US, Stanford researchers using payroll data from ADP report that employment of 22- to 25-year-olds in AI-exposed jobs is 19% below that of peers in less-exposed jobs, as of June 2026. Experienced workers show no such gap, and it comes mainly from companies hiring fewer young people.

For students aiming at office work, or anyone thinking of moving into it, I think the key is not to compete on "I can do office work" alone.

VIEW03

Office work is shifting from typing it in to checking it and connecting people

None of this means office work has no future.

AI drafts documents, enters data and totals tables quickly. But someone still has to check that two tables really match over the year they overlap, notice an odd step in the middle of a data series, and decide which numbers are fit to show other people. Even in making this article, the moments that needed a human were exactly those: checking, judging and explaining.

From a company's point of view, an office worker who can run AI and take responsibility for whether the documents and numbers are right is someone they want to keep. Office work is moving from doing the tasks by hand to checking AI's output and connecting one department with another. Rising pay for office work as a whole may be a sign of that.

Indeed Hiring Lab's Japan analysis found that among people searching for jobs with "AI" in their query, office work accounted for 7.3% of clicks in January to June 2026, the second-largest share outside tech jobs. People looking for office jobs where they use AI are already on the move.

SOURCES ── References

  1. Ministry of Health, Labour and Welfare (MHLW), 'Employment Referrals for General Workers, September 2026' (October 30, 2026) (in Japanese)
  2. MHLW, 'Employment Referrals for General Workers: Reference Tables, September 2026' (October 30, 2026) (in Japanese)
  3. MHLW, 'Employment Security Statistics, long-term series Table 21: Labor market indicators by occupation (JSOC 2009), April 2022 onward' (e-Stat, updated October 30, 2026) (in Japanese)
  4. MHLW, 'Employment Security Statistics, long-term series Table 21: Labor market indicators by occupation (MHLW classification, 2011 revision), March 2012 onward' (e-Stat, updated August 30, 2024) (in Japanese)
  5. MHLW Employment Security Bureau, 'On the occupational classification from April 2023' (May 30, 2023) (in Japanese)
  6. MHLW Employment Security Bureau, 'On figures from September 2021' (October 29, 2021) (in Japanese)
  7. Statistics Bureau of Japan, 'Labour Force Survey, Historical data 6: Employed persons by occupation' (e-Stat) (in Japanese)
  8. Association of Job Information of Japan, 'Job advertisement counts, September 2026' (October 23, 2026) (in Japanese)
  9. Association of Job Information of Japan, 'Job advertisement counts, August 2024 to August 2026' (in Japanese)
  10. Association of Job Information of Japan, 'Job advertisements in 2025' and 'Market size of job information services' (February 13, 2026) (in Japanese)
  11. Cabinet Office, 'Annual Report on the Japanese Economy and Public Finance 2026', Chapter 2, Section 2 (July 2026) (in Japanese)
  12. MHLW, 'White Paper on the Labour Economy 2026', Part II, Chapter 2 (September 29, 2026) (in Japanese)
  13. Reuters, 'Japanese companies' appetite for hiring slips as AI spreads' (September 11, 2025, republished by Newsweek Japan) (in Japanese)
  14. Federal Reserve Bank of Dallas (Samuel Dodini, Tucker Smith), 'Job postings show early signs of AI automation impact' (September 1, 2026)
  15. Federal Reserve Bank of New York (Richard Audoly, Miles Guerin, Giorgio Topa), 'Do Job Postings Show Early Labor-Market Effects of AI?' (May 14, 2026)
  16. Stanford Digital Economy Lab (Brynjolfsson, Chandar, Chen), 'Canaries in the Coal Mine?', August 2026 update (August 12, 2026)
  17. Indeed Hiring Lab Japan (Yusuke Aoki), 'AI-related job searches up 13-fold since ChatGPT' (July 14, 2026) (in Japanese)
  18. Kyoto Labour Bureau, 'Overview of the Kyoto Labour Bureau (Hello Work Day)' (in Japanese)
  19. MHLW Employment Security Bureau, 'Hello Work: main activities and results (April 2026)' (in Japanese)
  20. DATA TIMES Editorial Team, 'Hello Work openings for general clerks: e-Stat Table 21 spliced from March 2012 to September 2026, with method' (retrieved and calculated November 7, 2026)
#BUSINESS#Office jobs#Clerical work#Job openings#Hello Work#Jobs-to-applicants ratio#Japan labor market#Generative AI#AI and jobs