Wharton’s great contrarian says AI adoption isn’t an easy way to cut headcount: ‘The key thing … is just how much work is involved in doing it’

“You didn’t have to think very long to realize that just wasn’t going to make sense in practice,” Cappelli told Fortune on Zoom from his home in Philadelphia.

“You didn’t have to think very long about driverless trucks to think about, okay, what happens when they need gas? You know? Or what happens if they have to stop and make a delivery? And if they have to have an employee sitting with them, of course it defeats the purpose, right?”

Cappelli, who recently partnered with Accenture on a series of podcasts to get to the bottom of what AI is actually doing to jobs, warned against listening too closely to the companies that are talking their book, or trying to sell you on their new products.

“If you’re listening to the people who make the technology, they’re telling you what’s possible, and they’re not thinking about what is practical.”

Over the course of a wide-ranging conversation with Fortune, Cappelli tackled what AI is really doing to work, much like he talked to Fortune previously about how remote work is, actually, quite bad for most organizations.

“I mean, people say I’m a contrarian,” Cappelli said, “but I don’t think so, so much as I just am skeptical about stuff, you know?”

When pointed out this was an inherently contrarian position, Cappelli laughed, before returning to the main point. “I just get nervous with hype.”

He talked to Fortune about how his research fits into the wider picture that defined the back half of 2025, after the influential MIT study that caught the eye on 95% of generative AI pilots failing to generate any meaningful return. His favorite example was a particular case study on a company that actually made AI work, both cutting headcount and boosting productivity. It still didn’t fit neatly with predictions (say, from Elon Musk or Anthropic’s Dario Amodei, that work will soon be optional, or even a hobby). “It’s hugely expensive to do this,” Cappelli said about his findings. “And this was a success.”

Three times the cost

Cappelli detailed the findings of a case study that he participated in, published in the Harvard Business Review, on Ricoh, an insurance claims processor: the exact type of low-level administrative work that AI is supposed to automate easily. The reality of adoption, however, was a financial shock. While the company eventually achieved three times the performance, the transition was anything but cheap. The firm spent a year with a team of six, three of whom were expensive outside consultants, just to get the system running.

“The first thing they discovered,” Capelli said, “is large language models could do this pretty well — at three times the cost of their employees doing it [manually]. Okay, so that’s not going to work.” Cappelli pointed out that the costs included Ricoh paying roughly $500,000 in fees to outside consultants.

Even after optimizing the process, Ricoh was still spending about $200,000 a month on AI fees—more than their total payroll for the task had been. They were able to cut their headcount from 44 to 39, he added, showing just how far from being a massive job killer AI is in practice. His explanation recalls his self-driving truck example.

“The reason they still need employees is that lots of problems have to be chased down, and they’re harder to chase down if they come off of AI,” he said. The good news, he added, is that this Ricoh division will ultimately be three times as productive.

“So that’s the payoff, but it’s not cheap [and] it took a hell of a long time to do.”

Ashok Shenoy, VP of Ricoh USA, told Fortune that, after starting to use AI for “very routine, repetitive, high-volume tasks,” work for humans didn’t disappear, but “shifted toward areas where human judgment and experience add the most value.” In the year or so since the case study was conducted, he noted that Ricoh has successfully applied AI to mid-level, repetitive, time-consuming tasks at scale, and expects to use AI agents to achieve partial or full workflow automation within the next six to 12 months, “with a human-in-the-loop to resolve missing or unclear information and ensure quality.”

While acknowledging the big-ticket costs highlighted by Cappelli, Shenoy noted that this project reached break-even in less than a year, and it’s $200,000 monthly costs are less expensive than the previous operating model. “The shift to AI delivered an estimated 15% total cost reduction, even though it did not rely on significant labor cuts.” Regarding headcount, he said “this exercise was not driven by cost or headcount reduction,” and AI implementation requires creating new roles, redesigning existing ones, and repurposing team members toward higher-value work. He said there haven’t been further job cuts, either, with staffing levels largely stabilizing as productivity increased and volumes grew. “The bigger change was in how people spent their time. They are doing less repetitive work and are more focused on resolving exceptions, maintaining quality and serving customers.”

Performative AI shame in the boardroom

Cappelli said he found similar dynamics in his partnership with Accenture, which looked at Mastercard, Royal Bank of Scotland, and Jabil. “These are all success stories,” he said, and in the long run, they will see productivity will go up. Companies will be able to do more with fewer people but “it’ll take a long while to get there.” He argued that something crucial is being underestimated. “The key thing, though, is just how much work is involved in doing it.”

Also, regarding headcount reductions, Cappelli said that at least in the areas that he researched, which were specific units within each company, he didn’t see any job cuts whatsoever. When contacted for comment by Fortune, Accenture said it largely agrees with Cappelli’s conclusions, and referred back to CEO Julie Sweet’s recent interview with Fortune Editor-in-Chief Alyson Shontell.

According to Cappelli, so much of the noise around AI—and the distance between what’s possible and what’s practical—is driven by what other commentators have called “AI shame.”

Cappelli wasn’t familiar with the “AI shame” phrase, but told Fortune it was “absolutely right” in describing what he’s seen. “They’re pretending so they can say they’re doing something, right?” he said. “So the pressure is just enormous on them to try to make this stuff work, because the investors love the idea.”

The professor cited the Harris Poll’s finding in early 2025 that 74% of CEOs globally felt they’d lose their job in two years if they couldn’t demonstrate AI success, and roughly a third said they were performatively adopting AI without really understanding what it would entail. As The Harris Poll put it: “CEOs estimate that over a third (35%) of their AI initiatives amount to mere ‘AI washing’ for optics and reputation, but offering little to no real business value at all.”

Cappelli described how markets typically celebrate news of layoffs, and even cited research that “phantom layoffs” get announced by companies that never actually occur, because companies are arbitraging the positive stock-market reaction to the news of a potential layoff.

Cappelli predicted a “slow learning curve” will take place, in which CFOs will start realizing “this is super-expensive stuff to put in place.” The problem, according to Cappelli, is that U.S. management has become “spoiled” and increasingly averse to the hard work of organizational change.

“[Employers] think it should be free. It should be cheap. You should just be able to hang a shingle out, and the right people will just show up,” he says. Real AI success, in his opinion, will require “old-fashioned human resources” work: mapping workflows, breaking down jobs into tasks, and having employees work alongside AI “agents” to refine prompts.

“You can’t do it over the top of employees, because the employees really do know how their job is done,” Cappelli said. The professor was withering about what he sees happening in most C-suites, saying they are largely “ducking” the problem of really grappling with this technology.

“They’re not seeing it as an organization change problem and a big one,” he said. “They’re just stressing everybody out and, you know, hoping that it somehow works itself out.”