Thought Leadership
Aaron Levie hands his fellow executives an unflattering diagnosis: CEOs systematically overestimate AI because they sit too far from the operational work where it has to deliver. Studies from Berkeley, the NBER and MIT supply matching numbers.
When an AI advocate with 2.7 million followers on X tells his fellow CEOs they are suffering from “AI psychosis”, it is worth a closer look. Aaron Levie, co-founder and chief executive of cloud provider Box and an active angel investor in AI startups, made the accusation in a lengthy post over the weekend. The interesting part is not the label but the reasoning behind it. According to Levie, CEOs are structurally cut off from precisely those work steps where it gets decided whether an AI promise actually holds up.
Demo instead of deployment
Levie’s argument hinges on an image he calls the “happy path”. A chief executive has a prototype built or a contract generated, sees the clean result and concludes that AI agents could fully take over these tasks in the future. What gets ignored from the boardroom, he argues, is the chain of follow-up work: someone has to review the code before production and patch the bugs, identify calls to hallucinated libraries, strip legally risky clauses from the contract and wire everything up to existing agreements.
Levie’s recommendation to his peers is correspondingly pragmatic. Use AI yourself, intensively and over a long period, rather than relying on demo impressions. Only then does a realistic sense of potential and limits emerge. Those who actually do this, he suggests, are currently in the minority.
What the research says about Levie’s diagnosis
The thesis that AI’s value is overestimated can be backed up empirically. A meta-analysis in UC Berkeley’s California Management Review from October 2025 evaluated 371 estimates from the years 2019 to 2024 and found “no robust, publication bias adjusted relationship” between AI adoption and aggregate labor market effects.
A working paper published in March 2026 by the US National Bureau of Economic Research (NBER) reaches a more differentiated conclusion. The authors surveyed nearly 750 executives and did find positive productivity effects, but explicitly describe a “productivity paradox” in which perceived gains are larger than measured gains, presumably because revenue effects are delayed.
An MIT analysis from the Initiative on the Digital Economy, finally, projects that AI models will only reach success rates of 80 to 95 percent on most text related tasks around 2029, and that at a quality level the authors describe as “minimally sufficient”. Before agents broadly outperform humans, several additional years are likely to pass.
Layoffs justified with AI
While research is urging caution, layoffs in the tech industry are running at high speed. The tracker Layoffs.fyi counts 115,430 cut positions across 152 companies in the first five months of this year. For all of 2025, the same source recorded 124,636 employees affected at 275 firms. This year’s pace is markedly above last year’s.
Many corporations cite AI productivity gains as the justification. Critics counter with the term “AI washing”, meaning the retroactive AI label slapped onto cost cuts that were already on the agenda. ClickUp CEO Zeb Evans is unusually candid about it. He announced on X that he had laid off 22 percent of his workforce after the company internally deployed around 3,000 AI agents. The goal, he insists, is not cost reduction but an organization whose employees primarily review the outputs of these agents, a “100x organization”, as Evans calls it.
Market data fits the picture
That such architectures rarely run smoothly in practice is also suggested by a survey from S&P Global Market Intelligence. According to the firm, the share of companies abandoning the majority of their AI initiatives jumped from 17 to 42 percent within a year. On average, an organization scraps 46 percent of its AI proof of concepts before they reach production.
When everyone produces, the jam moves upward
Even when agents do their work satisfactorily, the efficiency equation does not automatically balance out. A study in the Harvard Business Review describes how the bottleneck shifts in many companies. As soon as the entire workforce produces more output with the help of AI, managers suddenly have to handle considerably more decisions, reviews and approvals than before. The bottleneck migrates from producing to approving.
Anyone who softens the approval step and instead creates autonomously acting agents risks the opposite outcome: organizational chaos. Levie’s own diagnosis can be summed up at the end without any clinical vocabulary. Executive suites are currently making personnel decisions affecting thousands of people based on an image of AI assembled from demo sessions and prototypes. Whether this turns out to be one of the rare business cycles in which reality lags behind the hype, or whether the operational evidence eventually does materialize, will not be decided on X. It will be decided by whether AI still works as a justification for the next layoff wave in 2027.
