When Klarna, the Swedish fintech giant, announced in 2024 that its AI chatbot was doing the work of 700 customer service agents, it sounded like a template for the future. Cut the humans, keep the software, watch the savings roll in. Dozens of major employers followed a similar script. By early 2025, a survey from the Upwork Research Institute found that roughly 80% of companies that had deployed AI tools went on to reduce headcount. The logic seemed airtight: if the machine can do the job, why pay a person?
But a growing stack of research suggests that logic is broken. A plant-level study from the National Bureau of Economic Research found that layoffs did not lead to subsequent productivity gains at the establishments examined. And a separate NBER field experiment showed that AI could dramatically boost worker output without eliminating a single role. Together, the findings point to an uncomfortable conclusion for executives still planning to swap headcount for algorithms: the layoffs may not be paying off.
The productivity gains that never arrived
The case against slash-and-deploy starts with hard numbers. The NBER plant-level study, authored by Martin Neil Baily, Eric J. Bartelsman, and John Haltiwanger, tracked manufacturing establishments before and after workforce reductions. Its core finding was blunt: layoffs did not produce measurable productivity improvements at the plants that carried them out. Workers left, but output per remaining worker did not climb to compensate.
That research predates the current AI wave, but its relevance has only sharpened. The pattern it documents keeps repeating across economic cycles. Companies announce cuts, promise shareholders that a leaner operation will be a more profitable one, and then struggle to realize those gains. What the study captures is a structural problem with layoffs as a productivity strategy, not a one-time anomaly.
More recent evidence reinforces the point. Klarna’s own 2025 earnings reports showed that while headcount fell, the company’s path to profitability remained bumpy, with customer satisfaction scores dipping and rehiring in some areas quietly resuming. IBM, which announced in 2023 that it would pause hiring for roles AI could fill, has not published data showing that the move improved per-employee output. The announcements were dramatic; the follow-through has been harder to measure.
What happened when companies kept their workers
The contrast comes from a paper titled “Generative AI at Work,” authored by Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond and published as NBER Working Paper No. 31161. The researchers studied what happened when a generative AI assistant was deployed alongside 5,179 customer support agents at a large software company. Nobody was fired. The AI sat in the workflow as a real-time coaching tool.
The results were striking. Agents using the AI resolved 14% more issues per hour on average. But the gains were not evenly distributed. Newer, less experienced agents saw the largest improvements, while top performers saw smaller bumps. In effect, the AI acted as a leveler, pulling the least skilled workers closer to the performance ceiling of their most seasoned colleagues.
“The AI tool essentially disseminated the tacit knowledge of the best workers to everyone else,” Brynjolfsson noted in discussing the findings. That is a fundamentally different model from the one driving most corporate AI strategies. Instead of replacing the workforce, the technology made the existing workforce substantially better.
Why the 80% figure deserves scrutiny
The Upwork survey that produced the 80% statistic is useful as a directional signal, but it comes with caveats. Surveys capture what respondents report, not independently verified operational data. What one company calls “AI deployment” might mean a full-scale automation overhaul; at another, it might mean a pilot chatbot in one department. The time horizon between adoption and layoffs also varies wildly. A company that cut 5% of staff six months after an AI rollout and a company that eliminated an entire division the same week are both counted the same way.
Response bias matters too. Companies that deployed AI and did not cut jobs may have been less likely to participate in a survey framed around workforce disruption. The 80% figure almost certainly overstates the rate at which AI directly caused layoffs, even if it accurately reflects a broad corporate impulse to pair new technology with headcount reductions.
None of this means the trend is imaginary. High-profile examples are easy to find. Chegg, the education company, saw its stock collapse after ChatGPT launched and subsequently cut roughly 20% of its workforce. Duolingo reduced contract workers in favor of AI-generated content. UPS announced 12,000 job cuts in early 2024, citing automation and efficiency. The pattern is real. The question is whether it is working.
The data gap executives should worry about
One of the most striking absences in the current debate is firm-level financial data linking AI deployment dates to post-layoff operating margins. Companies rarely disclose enough detail to isolate the effect of a technology rollout from other restructuring moves happening simultaneously. A firm might deploy AI, cut 10% of staff, renegotiate supplier contracts, and close two offices in the same quarter. The earnings call credits AI. The 10-K filing buries the specifics.
Longitudinal data on displaced workers is equally thin. When a company eliminates roles after adopting AI, some of those workers find new positions internally, some leave the industry, and some exit the labor force entirely. The Bureau of Labor Statistics tracks broad displacement through its Displaced Workers Survey, but the data is not granular enough to isolate AI-driven separations from other causes. Without that clarity, the full economic cost of these decisions stays partially hidden from policymakers and from the companies making them.
The customer support study, while rigorous, examined one occupation at one company. Whether its results generalize to manufacturing floors, financial trading desks, logistics networks, or creative studios remains an open question. Productivity gains for agents handling text-based conversations may not translate to environments where physical tasks, regulatory constraints, or creative judgment dominate the work.
What the research actually suggests companies should do
The practical signal from the available evidence is not subtle. Organizations that used AI to augment existing workers measured real, replicable productivity improvements. Organizations that used workforce reductions as a blunt cost lever, historically, did not see productivity follow. The research does not prove that no layoff ever pays off, but it directly challenges the default assumption that fewer workers plus new technology equals better performance.
The most defensible approach, based on what the data supports as of mid-2026, is sequential rather than simultaneous. Deploy the technology first. Measure its actual effect on output over multiple quarters. Then make staffing decisions grounded in observed results rather than projected savings. That sequence is slower and less dramatic on an earnings call, but the evidence suggests it is far more likely to produce the gains that executives keep promising investors.
For workers caught in the middle of these decisions, the takeaway is more immediate. The companies most likely to get AI adoption right are the ones treating the technology as a tool for the people they already employ, not as a justification for showing them the door. The research record, thin as it still is in places, consistently points in that direction. The companies ignoring it are making a bet that the numbers, so far, do not support.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
