AI agents cut task time by 87%, Harvard and Perplexity study finds

A new study from Harvard Business School and Perplexity takes a real-world look at what happens when you let an AI agent take the wheel. The research analyzes actual usage data from Perplexity's agent, aptly named Computer, and compares it to using the company's standard search product. The headline numbers are striking: after handing work off to the large language model, the average time to complete a task dropped by 87%, and the average labor cost fell by 94%. Those estimates come from a time-cost model the researchers built, then plugged into wage data from the Bureau of Labor Statistics (as of May 2025). The team calls the figures "approximate," but they're still eye-popping.

Digging into the mechanics reveals why. A search product forces you to manually walk through every step yourself. Computer, on the other hand, runs entire workflows on its own — the average session clocks 26 minutes of machine time, compared to just 33 seconds for search. The human's role shifts from assembly-line operator to manager: you set the goal and supervise the output. Not only do users submit nearly half as many "unsatisfied" follow-ups, they're also more willing to tackle complex, non-routine tasks outside their usual expertise.