The AI Productivity Paradox

Summary

William Roush examines the gap between individual reports of AI productivity gains and broader executive disappointment. He cites a Fortune report saying 90% of executives said AI had not boosted productivity, along with his claim that 95% of enterprise AI pilots return nothing and that companies are reconsidering AI spending. In his own work, he reports results ranging from being two to three times slower to 20% to 500% faster, depending on the task: basic, code-heavy interface work improves substantially, while complex work narrows the gap or becomes slower because LLMs make errors, choose poor approaches, and require guidance. He argues that the decisive factor is not simply using an LLM, but knowing when to stop using it and when to direct it toward an existing library or a narrower change. Roush says his company therefore treats AI use as a quality and coordination issue, aiming to prevent customers from receiving low-quality output and teammates from inheriting extra work. He criticizes “human-to-LLM interfaces” who pass questions and generated replies between colleagues without adding judgment, saying this creates longer communication, awkward output, and engineers who may not understand their own work. The article’s broader argument is that LLMs can expand the blast radius of unskilled and overconfident workers by making them faster at producing changes others must review. Roush describes failures in which Claude proposed unnecessary state management, caused a staging environment outage, and misdirected debugging until a human intervened. He concludes that delegation only creates value when someone evaluates the result, and recommends enforcing this expectation through workplace performance policies, including possible termination after an unsuccessful performance-improvement process.