Box CEO warns of 'AI psychosis' as tech layoffs outpace productivity gains

Box founder Aaron Levie has characterised a growing disconnect among technology leaders as "AI psychosis," a condition stemming from executives' distance from the practical "last mile" of work required to generate value. Levie argues that while CEOs experiment with AI prototypes, they often fail to grasp the complex human oversight needed for tasks such as reviewing code, identifying bugs, or training models on idiosyncratic corporate data. This commentary highlights a broader tension in the industry, where enthusiasm for automation clashes with the operational realities of deployment.

The assertion comes against a backdrop of significant workforce reductions in the technology sector. Data from layoff tracker Layoffs.fyi indicates that nearly 115,000 people were laid off across 152 companies in the first five months of 2026. This figure approaches the total workforce reductions seen in all of 2025, with many organisations citing AI efficiency as a primary driver for these cuts. However, critics argue that some firms are engaging in "AI washing," attributing cost-saving measures to artificial intelligence when other business metrics may be the true cause.

Specific corporate strategies illustrate the scale of this shift. ClickUp CEO Zeb Evans recently laid off 22 per cent of his workforce after deploying approximately 3,000 AI agents for internal tasks. Evans described the move not as a cost-cutting exercise, but as an effort to create a "100x org" where human employees focus on reviewing agent output rather than performing the work themselves. Levie, an active angel investor in AI startups, advises leaders to engage deeply with the technology to understand its limitations alongside its potential.

Despite these corporate ambitions, academic research challenges the narrative of immediate productivity boosts. A meta-analysis published in UC Berkeley’s California Management Review found no robust relationship between AI adoption and aggregate productivity gains. Similarly, research from the National Bureau of Economic Research noted a "productivity paradox," where perceived gains significantly exceed measured outcomes, suggesting that the integration of AI has not yet delivered the promised efficiency dividends.

Further studies from MIT and the Harvard Business Review warn of structural risks. MIT researchers project that while large language models may achieve 80 to 95 per cent success rates on text-related tasks by 2029, agents will require several more years to consistently outperform humans. Concurrently, Harvard Business Review research indicates that widespread AI adoption often shifts bottlenecks to executive levels, as leaders must authorise the increased volume of output. Without adequate oversight, Levie warns that this dynamic could lead to organisational chaos rather than streamlined operations.