A founder in a spare bedroom in Columbus spends a weekend building a tool that sorts insurance claims by how likely they are to be disputed. She used to do the sorting herself, for an adjuster’s salary. In March a regional insurer pays her by card for the first month. She has no employees, no office lease and no payroll account, so when the statistician at the state labour office tallies new jobs that quarter, she is not in the count. Her old employer, meanwhile, has stopped replacing the two adjusters who left.

The scenes that open and close this article are illustrations; the people in them are invented.

Whether the founder and the empty desks cancel out is the question of a preprint posted to arXiv on 12 November 2026. Mark Esposito and Liu Zhang, both listed with the company micro1, argue in “AI and the Two Margins of Work” that technology acts on labour demand in two places. One is work inside established firms, where automation can remove tasks quickly. The other is the creation of new activity, often by new firms, which takes years to turn into hired workers. Payroll statistics, they write, record the first promptly and the second late. The paper has not been peer reviewed.

What computing did to the two margins

The authors start with the record of the last big general-purpose technology. Using the Census Bureau’s Business Dynamics Statistics, they look at 1982 to 1998. Firms in their first year added an average of 2.78 million jobs a year. Young incumbents, aged one to five, lost a net 0.44 million, and older incumbents lost 0.52 million. The published national average was 1.80 million new jobs a year, so entry more than covered the shrinking of everyone else.

Those new jobs did not vanish. Employment in firms that entered in 1978 or later rose from 2.55 million in 1978 to 45.68 million in 1998, while employment in the firms already operating before 1978 fell. By 1998 the later entrants held 42 percent of payroll, and an entering cohort kept about 80 percent of its payroll after five years. Across industries, those most exposed to automation saw incumbents fall behind from the early 1980s. The authors call the later entrant build-up only suggestive, because its standard errors exceed the estimates by 1998.

They set this against work by David Autor and colleagues, who compared decades of Census job titles and estimated that roughly 60 percent of US employment in 2018 was in titles that did not exist in 1940. Daron Acemoglu and Pascual Restrepo supply the frame: a technology has a displacement effect, as machines take over tasks, and a reinstatement effect, as new tasks appear in which people have an advantage. The long-run result depends on the balance.

A model of the lag

The paper’s model has one AI capability that lowers the cost of acceptable output in existing production and also makes additional projects worth trying. It keeps a founder’s own effort separate from the hours the founder pays for. The central assumption is that operating products take time to re-scale. If they re-scale at once, automation displaces no aggregate hours. If they do not, hours fall first and recover as incumbents expand again and new products arrive.

Timing then decides how deep the dip is. A sudden efficiency gain of the size the authors assume cuts hired hours by 4.6 percent. The same gain spread over seven years, which the authors say is closer to how AI has been adopted, produces a trough of 0.23 percent that closes after about seven years. These are simulations from calibrated assumptions, not measurements.

What the data show so far

The authors then look at establishment job flows from the Bureau of Labor Statistics through the first quarter of 2026. Across 72 industries, employment at establishments that existed before ChatGPT fell by 13.6 percent of its 2019 level between the fourth quarter of 2022 and the first quarter of 2026, at a pace no faster than in 2019 to 2022, and openings more than replaced it. Industries more exposed to large language models show no larger losses at continuing establishments than their peers; the estimated difference is 1.1 points with a standard error of 1.9. They also show no more job gains at openings.

A sudden shock of the size the model considers is ruled out. What remains cannot be sorted. The result, the authors write, is what their model predicts under slow adoption, and it is also what no effect on payroll would look like. They say what would separate the two: a difference should appear first within occupations and in the hiring of entrants to exposed jobs, before it shows up in industry totals.

The Bank of England’s November Monetary Policy Report, published on 6 November, reads the same shape from a different angle. Firms in its Decision Maker Panel estimate that AI has raised their productivity by about 0.5 percent over three years, which the Bank converts to 0.1 to 0.2 percentage points a year since early 2024, and report close to zero effect on employment. The Bank’s Agents say adoption is reducing the likelihood of hiring or replacing workers, but not materially increasing firing.

Where the new work hides

The creation margin shows up earlier, but not in payroll. Marchesi and Tang estimate that after ChatGPT’s release firm formation rose by more than 10 percent in industries whose startup tasks AI can help with. Bena, Bian and Giannetti find about 20 percent more startup formation in industries with higher task-level exposure, with smaller entrants. Both studies count applications and registrations, which come before anyone is hired.

The authors add a measure of what payroll misses. Census counts of businesses without employees rose 5.1 percent in 2024, or 1.56 million, against a net private payroll gain of 0.89 million. Valued at average receipts and converted at average pay, they put the increase at about 1.3 million payroll-job equivalents, and they warn that most such businesses are part-time and that the counts alone do not tie the rise to AI.

The bigger claim concerns the record. If hours in new activities enter the statistics with a delay of one to three years on average, the 0.23 percent true trough is measured at 0.6 to 1.0 percent, and the measured recovery moves from 6.9 years to between 7.7 and 11.7. With no conversion of new activity into payroll, the measured trough reaches 2 percent. The underlying path is the same, they write; only the record is late.

A ratio to watch

Their proposed fix is a net reinstatement ratio: paid hours in additional activities divided by paid task hours displaced since a base year. Above one, new work uses more hours than automation removed. Displaced hours need a counterfactual, since nobody observes the work that would otherwise have been done. As a historical analogue, the authors compute 1.09 from the Autor titles for 1990 to 2018 and stress that this is not an estimate for AI.

There are limits the authors name themselves. Job titles lag reality, a lawyer can keep a title while the work changes, activity can sit outside payroll, and an early observation describes one stage. The model’s calibration drives its dip and recovery, and the data window includes the pandemic recovery.

By autumn the founder in Columbus posts her first opening: a part-time analyst to read the disputed claims she cannot get to. It will appear in a payroll survey two quarters later. The insurer that stopped replacing adjusters will by then have a year of figures showing it needs fewer of them, and the statistic that counts both will have only one side filled in.

References

  • Acemoglu, D., & Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives, 33(2), 3-30. doi.org
  • Autor, D., Chin, C., Salomons, A., & Seegmiller, B. (2024). New Frontiers: The Origins and Content of New Work, 1940-2018. Quarterly Journal of Economics. doi.org
  • Bank of England (2026). Monetary Policy Report, November 2026. bankofengland.co.uk
  • Bena, J., Bian, B., & Giannetti, M. (2026). Prompted to Start: How Generative AI Is Transforming Entrepreneurship. SSRN 5749564. papers.ssrn.com
  • Esposito, M., & Zhang, L. (2026). AI and the Two Margins of Work: Why Job Losses Show Up Before Job Creation. arXiv:2611.17530. arxiv.org
  • Marchesi, E., & Tang, S. (2026). When AI Meets Entrepreneurship: Evidence from the Commercialization of ChatGPT. SSRN 5759264. papers.ssrn.com
  • U.S. Census Bureau. Business Dynamics Statistics. census.gov