AI Agents Are Making Knowledge Workers Busier, Not Freer

This position paper argues that AI agents, despite promising to free up human time, are actually making workers busier, not more rested. We refine this claim into a testable sub-claim: AI increases the amount of work items (pull requests, lines of code, meetings, emails, messages, draft documents) without a corresponding rise in genuine output. Humans then manage the overflow through review, supervision, and coordination work. We see this as an example of the Jevons Paradox, which is a 19th-century observation that efficiency gains in resource usage often lead to increased total consumption. For instance, more efficient steam engines led to burning more coal, and washing machines resulted in higher cleanliness standards instead of more leisure. Similarly, AI agents are advancing the limits of what can be achieved in a workday faster than they are lessening the effort needed for current tasks. We draw on evidence from labor economics, organizational psychology, cognitive science, and deployment studies across three areas: software engineering, academic research, and organizational work. This evidence shows that this trend is a fundamental aspect of how productivity tools interact with competitive labor markets and human behavior. We urge the machine learning community to investigate the second-order effects of deploying AI agents. They should develop metrics for human well-being alongside task output and consider the “net effect on human workload” as a key evaluation measure for agent systems.