Researchers from Harvard Business School and Perplexity released a paper on June 5 measuring what changes when people delegate work to an AI agent instead of running searches themselves. They compared Perplexity's Computer agent against its Search product using production data, and the headline number is a task that took 269 minutes with Search dropping to 36 with the agent. That is the kind of figure a company likes to publish about its own product, so it is worth poking at.
The matched-pair trick
The interesting part is the method, not the press-release stat. The authors found 10,000 session pairs where the same user typed near-identical opening queries (cosine similarity above 0.99) into both Search and Computer, then treated those as a rough natural experiment for the same underlying task done two ways. The agent ran 26 minutes of autonomous work per session. Search ran 33 seconds. That 48-fold gap in machine time is the engine behind every other number in the paper.
One catch they are upfront about: they never observe the actual human time a Search user would have spent. The 269-minute counterfactual is estimated, three different ways (a tool-based tally, an LLM guess from query text alone, and interviews with 25 users). When the load-bearing comparison is a number nobody directly measured, treat the precision with some suspicion. The direction is probably right. The decimal places, less so.
Quality went up, which is the surprising bit
You would expect faster-and-cheaper to come with a quality tax. The data says otherwise, at least on their proxy. Computer drew medium-to-high dissatisfaction on the next user turn 1.3% of the time versus 2.9% for Search.
A 55% reduction in dissatisfaction, the authors report, though "dissatisfaction inferred from your follow-up message" is a thin proxy for whether the work was actually good. Someone who gives up is also someone who stops complaining.
What people actually did with it
The scope findings are where this gets more useful than the speed pitch. Computer queries crossed outside the user's main occupation more often than their Search queries did, with a 9 percentage point average gap holding across all eight occupation clusters the team looked at. Create-level work, the top of the cognitive ladder, made up 50% of Computer queries against 26% for Search.
Then the number that earns the paper its thesis: 23% of Computer queries involved a task the same user had never once sent to Search. Software, documents, data visualization. Work that apparently was not worth doing when it meant orchestrating every step by hand. Each agent query also leaned on 2.40 distinct knowledge domains on average versus 1.74 for Search, which the authors read as one person doing what used to need a small team.
So the bottleneck was never information. Search solved that years ago. The bottleneck was execution, and that is the thing getting removed. The practical read for anyone managing knowledge workers: the scarce skill shifts from doing the work to scoping it well and checking the output, because the agent carries a higher fixed cost (you have to specify and verify) against a lower per-step cost. Short questions stay in the chat box. Long multi-step jobs migrate.
Worth remembering whose product this is. Perplexity has every reason to show Computer in a flattering light, and the matched-pair design, clever as it is, can't fully rule out that people bring different tasks to an agent than to a search bar even when the opening words match. The paper, posted to arXiv on June 5 and not yet peer reviewed, is a v1 preprint. Expect the numbers to move.
