The clock is ticking for AI to deliver on its promises of transforming workplace and economic productivity, and if the lag in investment returns continues to lag, markets are in for a rude awakening, according to a top economist.
Torsten Slok, the influential chief economist at Apollo Global Management, argued in a recent blog post that there is a growing gap in AI-powered productivity. Basically, you only see this in technology companies, but not in most Fortune 500 companies.
While some sectors such as software and technology can easily integrate AI into their operations, Slok argued that adoption of this technology is slow in the vast majority of the economy. Due to regulatory hurdles, data protection and workflow integration, this requires time and effort, meaning structural productivity gains are slow and return on investment remains to be seen. Slok said he believes it could happen — eventually. And by that point, the stock bubble may have already burst, as the market has priced in the return sooner rather than later.
“The main issue is the length of the ROI runway outside of the tech sector,” Slok said. “The bottom line is that a mismatch between current profit expectations and the actual time it will take for companies to achieve ROI on AI investments could have a significant impact on the valuations of many AI companies today.”
Slok cites Bloomberg and Macrobond data indicating that profit margins for the Magnificent Seven rose from around 15% to 25% between the first quarters of 2023 and 2026, but profit margins for the rest of the S&P 493 were around 10%. The Bloomberg 500 Index follows the same pattern as the S&P and maintains a constant profit margin of 12% over the same period.
What worries Slok most is what happens if that gap widens while AI adoption and productivity gains continue to stall. A groundbreaking and controversial MIT study published last year found that only 5% of companies achieved a significant return on investment from generative AI pilots. The Apollo economist warned that markets will face a “painful repricing” that threatens to slow the AI boom as expected profits or current market prices continue to exceed actual profits.
“In other words, companies will curb their AI spending if they don’t see a quick return on investment,” he said.
Table of Contents
Where does the economy see the variable returns on investment in AI?
US industrial giants are already facing setbacks in their mass automation efforts. In a visible sign that human expertise is needed to truly leverage AI productivity improvements, Ford hired 350 “Graybeard” engineers—industry veterans, including former employees—to train junior staff and reprogram ineffective AI tools. The automaker continues to use AI vision systems at 33 plants worldwide, with more than 1,000 cameras performing millions of inspections on the assembly line, but has recognized that the technology is not as effective without human oversight.
“Artificial intelligence is a fantastic tool, but it is only as good as the information you train it with,” Charles Poon, Ford’s vice president of vehicle hardware engineering, told reporters last month. “In recent years, we have not paid as much attention as we should to the experience of our most knowledgeable engineers, who have accompanied us through many product cycles.”
Ford is following the lead of companies like IBM, which cut thousands of jobs last year due to increased spending on cloud services. In March, the company announced it would triple U.S. hiring across all divisions, arguing that more AI-focused jobs are needed.
As things stand, this human labor is far cheaper than the automation tools companies are investing in, further calling into question the productivity benefits of AI in the workplace. Nvidia’s vice president of applied deep learning, Bryan Catanzaro, said earlier this year that the cost of AI still far exceeds that of human labor, an admission that coincided with an era of tokenmaxxing in which tech companies like Meta incentivized AI use through internal employee leaderboards, leading to workers using the technology solely for its own sake, which simultaneously drove up costs.
According to Slok, the race to use tokens most effectively is more of an indication that companies are struggling to get value from AI and are not seeing real workplace gains from it.
“Companies will scale back their AI spending if they don’t see ROI quickly,” Slok said. “And the current focus on token optimization is an early warning that AI implementation may be a bumpier and slower road than expected.” (Slok has separately argued that AI will create more jobs, not fewer, as he has become Wall Street’s chief advocate for the relevance of the Jevons Paradox; he also believes it will lead to a boom in small business entrepreneurship.)
Why is AI not yet keeping its promises?
Peter Cappelli, a professor of management at the Wharton School of the University of Pennsylvania, recognized the problem highlighted by Slok early on and led a case study of Ricoh, a digital services company, published in the Harvard Business Review. In short, he found that people greatly underestimate “how much work is involved” in achieving productivity and ROI gains, as he said Assets Earlier this year.
“If you listen to the people who are developing the technology,” Cappelli said of the AI course, “they tell you what’s possible and don’t think about what’s practical.”
The gap between the possible and practical uses of AI is caused by “AI shame,” or pressure on the effectiveness of these new technologies, especially in the face of increasing pressure from investors. It’s a phenomenon seen at tokenmaxxing technology companies where leadership has mandated increased AI use but failed to articulate concrete use cases or goals related to AI use.
The Boston Consulting Group found in a recent study that using AI for the sole purpose of AI can actually negate the technology’s productivity gains. The consulting firm’s 2026 Global AI at Work report, which surveyed approximately 12,000 frontline workers, found that 42% of respondents reported saving eight hours of time per week through regular use of AI. However, most said they received little or no guidance on how to use the time saved.
“This whole tokenmaxxing thing has probably run its course, and now it’s putting a pretty big strain on their cost base,” said David Martin, global head of BCG’s People & Organization practice Assets. “A lot of companies have just made AI available to everyone, regardless of position, and I think now they’re going to say, ‘Let’s think more about who has access and what the business case is? And do we ultimately deliver that?'”
In Ricoh’s case, Cappelli said, when the company outsourced low-level administrative work of processing insurance claims to AI, the process required about $500,000 in outside consultant fees as well as $200,000 per month in AI fees, ultimately resulting in costs three times higher than if an employee were to do the administrative work manually. The company has reduced its workforce only slightly, from 44 to 39, Capelli said.
Ultimately, Ricoh increased the department’s productivity threefold, but it took time, and his example underscores Slok’s concern about what AI has to offer: productivity gains are possible, but they do not come without immense initial costs in terms of time and money.
“So that’s the payoff,” Capelli said. “But it’s not cheap [and] It took a hell of a long time.”
https://fortune.com/2026/07/06/ai-productivity-gains-bubble-painful-repricing-markets-torsten-slok/
