The office in this scene is made up. It stands for the kind of mid-sized freight forwarder where an AI licence arrives on a Monday and the rest of the week goes on as before. The operations director has bought seats for her thirty staff, sent one email with a link, and scheduled no training. By Friday she has watched what they do with it. A clerk pastes in a customs code and asks what it means. Another asks for a polite version of a late-payment reminder. Somebody uses it to look up a port. Nobody has changed how a shipment is booked. In the quarterly meeting the owner asks what the tool has saved, and she has no number to give him. He shrugs and says they should not worry too much about it yet.

Her week is invented, but the pattern it shows was described this month by two Harvard Business School professors. On 20 October HBS Working Knowledge republished an interview first printed by the Harvard Gazette on 29 September, in which Joseph Fuller and Raffaella Sadun, with the Kennedy School economist Douglas Elmendorf, try to explain a puzzle. Layoffs blamed on AI are in the news and forecasts are grim, yet the job market still looks ordinary. Their answers are partly about the technology and mostly about the people deciding how to use it.

Grim forecasts, ordinary numbers

The Gazette lays out the contradiction in a few lines. News outlets, it says, report more layoffs attributed to AI this year than last. Dario Amodei, the chief executive of Anthropic, told Axios in 2025 that AI could eliminate half of all entry-level white-collar jobs within five years and push unemployment to between 10 and 20 percent, a warning Fortune reported at the time. Entry-level postings have fallen by up to 73 percent in four years, by some counts.

Other figures point the other way. The unemployment rate is hovering around 4.1 percent, which the Gazette says is largely buoyed by growth in healthcare. For every article announcing AI-related layoffs, it adds, another dismisses the claim as “AI-washing”. And Accenture says 52 percent of top executives expect AI to increase entry-level hiring, not decrease it. The Gazette gives no way to settle which set of numbers is right, and it does not try to.

A third of the experiments

Fuller’s explanation starts with what firms do once they have the tool. With Accenture Research he built a model that analysed work tasks across all job categories. As of today, he says, 41 percent of all work tasks can be automated or augmented by AI. Firms are not rushing to do it. Only about one-third of their AI experiments succeed, he says, possibly because of faulty data inputs or employees who do not know what they are working with. The article does not describe how the task model was built, so the 41 percent is Fuller’s figure and Accenture’s, not an independently checked one.

“Companies are not training people formally, so employees are basically using it like glorified Google, and they’re not getting much productivity out of it,” Fuller said. A boss who sees little gain may conclude, in his words, “this isn’t as good as I was hoping, I guess we shouldn’t worry too much about this.”

He does not expect that to last. Adoption will rise, he warned, as cheaper, more specialised AI agents arrive. On this reading the quiet in the job market measures how badly companies are using the tools, which can change without any new breakthrough.

Scenarios, not forecasts

Elmendorf is wary of reading anything from today’s data. “What we’ve seen so far in the labor market from artificial intelligence has very little predictive power for what we’re going to see in the labor market because of AI in five years, or 10 years, or 15 years,” he said.

With Karen Dynan of the Kennedy School and Louise Sheiner of the Brookings Institution, he wrote a working paper for the National Bureau of Economic Research, issued in July, that sets out four scenarios. The first has AI raising productivity growth and income gains shared broadly. The second keeps the productivity gain but sends most of the income to the top earners. The third adds labour-market disruption, with displaced workers taking time to find new jobs and some leaving the workforce. In the fourth, productivity growth and job loss both peak, the losses are more permanent, and the additional income goes to the owners of capital. The authors stress that these are illustrative, not predictions, and the Gazette says they declined to say which is likelier.

The harshest case has about 3 million people, roughly 2 percent of the labour force, out of work at any given time. The authors compare that with the China shock of the early 2000s, which, by the estimates they cite, cost about 1.5 to 2 million jobs between 2000 and 2007. David Autor, David Dorn and Gordon Hanson, who studied that shock, found that adjustment in the hit local labour markets was “remarkably slow”, with wages and participation depressed and unemployment elevated for at least a decade. Dynan points out one difference in an interview with Harvard’s Griffin graduate school. The China shock hit concentrated manufacturing towns, whereas AI is more likely to reach people spread across industries, occupations and regions: call-centre workers and software coders today, administrative workers and lawyers tomorrow.

Elmendorf, for his part, thinks it very likely that there will be significant job losses in the coming couple of decades, but that most of those people will find new work. Dynan adds a result that sounds like good news and carries a catch. Under current law and with no policy response, debt in all four scenarios ends lower than the roughly 175 percent of GDP that the Congressional Budget Office projects, at about 125 to 135 percent. She notes that history suggests policymakers do not always let such automatic changes play out.

Who decides

Sadun puts the decision on executives. She works at HBS’s Digital Reskilling Lab, which retrains workers for AI-augmented roles and then tests whether the training worked. The technology, she says, gives access to expertise that was once dispersed, and executives wonder whether it can automate away the need to invest in human capital. “I think that’s a tremendous trap, a terrible trap,” she said. Employees hold tacit knowledge of their market, role and resources, which makes them more valuable than their job descriptions suggest.

Fuller describes one practical response. Executives tell him a 22-year-old with a new computer science degree may not offer the same value, but they also see that a company cannot have a senior engineer without first having a junior one. At one large tech company he declined to name, he helped them study senior engineers’ performance reviews and found that many lacked sales and product-management experience. The junior role was rewritten to keep some software engineering and add regular, structured exposure to both. The entry-level job, he said, may now look like a second or third job would have looked five years ago.

Elmendorf doubts that appeals of this kind can outweigh the profit motive. Companies will focus on what is good for them, he said, and for some that will mean AI instead of workers for some tasks. That is the case for government, which exists to solve collective-action problems that markets are not set up to handle. The paper’s options include wage insurance that pays laid-off workers part of the gap between old and new pay, subsidised private employment, shorter work weeks and expanded retraining. The authors weigh waiting for clearer evidence against a modest “insurance” approach now, such as a scaled-down trade adjustment assistance programme, that could be expanded if the harsher scenarios arrive. “Policymakers should be prepared for much bigger economic changes than we’ve seen so far from AI,” Elmendorf said. “Not because we’re sure those changes will happen, but for the same reason you buy insurance against possible problems in your personal life: It might happen, and you’d rather not be surprised when it does.”

The operations director of the invented forwarder goes back to her desk after the meeting. The tool is still open in thirty browsers and the reminder is still polite. Whether the owner’s shrug becomes a decision about her clerks depends less on the software than on whether anyone ever teaches them to use it, and on whether the owner asks that question before or after the next price list arrives.

References

  • Autor, D. H., Dorn, D., & Hanson, G. H. (2016). The China Shock: Learning from Labor Market Adjustment to Large Changes in Trade. NBER Working Paper 21906. nber.org
  • Boles, S. (2026). AI Hasn’t Spurred Widespread Layoffs Yet. Why Not? Harvard Gazette, republished by HBS Working Knowledge. library.hbs.edu
  • Boles, S. (2026). Why AI hasn’t triggered mass layoffs, yet. Harvard Gazette. news.harvard.edu
  • Dynan, K., Elmendorf, D., & Sheiner, L. (2026). How Might Fiscal Policy Respond to the Rise of Artificial Intelligence? NBER Working Paper 35437. nber.org
  • Harvard Kennedy School (2026). Fiscal policy will need new tools for uncertain AI future. hks.harvard.edu
  • Massari, P. (2026). How AI Might Impact the Economy and What Government Could Do About It (interview with Karen Dynan). Harvard Griffin GSAS. gsas.harvard.edu
  • Morris, C. (2025). Anthropic CEO warns AI could eliminate half of all entry-level white-collar jobs. Fortune. fortune.com