Perplexity published efficiency numbers for its Computer agent product on June 8, co-authored with Harvard Business School researchers, and the headline figures are already circulating: an 87% reduction in task time and a 94% reduction in task cost compared to a search-plus-human baseline. Those numbers will appear in vendor briefings and purchasing justifications across enterprise teams by Q3. Before they do, they deserve a careful read.
The methodology is documented and defensible at a high level. Perplexity matched Computer sessions against Search sessions with near-identical queries across 10,000 pairs, then estimated the human execution time that Search would require by classifying tool calls into “search” actions (already handled by the product) and “do” actions (steps a human would perform manually afterward). Hourly wage data from the U.S. Bureau of Labor Statistics set the cost baseline. The 87% figure is a tool-based estimate; an independent LLM-based estimate arrives at 84%, and 25 user interviews produced a median speedup of 25x with a wide range.
The task mix matters here, and Perplexity acknowledges it. Research and Analysis accounted for 25.8% of sampled Computer queries, and Document and Asset Creation accounted for another 18.6%. Together, nearly half the measured workload was synthesis and production work. These tasks share a structural property: they have reasonably clear terminal states, their inputs are largely parallelisable, and they do not depend on coordination with other humans or on judgment calls that require institutional context. Autonomous execution performs well against exactly that profile.
Contrast that with the DX Research findings on software engineering teams, which measured roughly 8% productivity gains across broad engineering workflows. Both numbers can be accurate. The difference is task structure. Software development involves continuous feedback loops, review gates, shared context across codebases, and ambiguous correctness criteria. A research-and-synthesize task concludes when the document is done; a production deployment concludes when stakeholders agree it is ready. Those are different problem classes, and agent productivity compounds differently in each.
The cost number carries an additional caveat that Perplexity does not foreground: the 94% reduction combines model compute costs at current Perplexity subscription pricing (which reflects subsidised rates at this stage of the product’s growth) with BLS wage data. As compute pricing normalises and as agent products move toward commercial rather than growth-stage pricing, the cost gap will narrow. The 94% figure is not wrong, but it is a current-pricing snapshot, not a stable structural claim.
What Perplexity does show convincingly is a behavioral shift. Computer users work outside their primary occupation cluster 59% of the time, versus 50% for Search users. Computer queries score higher on Bloom’s Revised Taxonomy, with 76% requiring higher-order cognition versus 55% for Search, and 50% classified as Create-level tasks versus 26% for Search. These are usage-pattern findings drawn from Perplexity’s own logs, not extrapolated claims, and they describe something real: when execution is automated, people attempt different work, not just the same work faster.
The honest procurement read is this. For teams whose core knowledge workflows are research-heavy, multi-source synthesis tasks, the efficiency gains are plausible and worth piloting with real task samples. For teams whose workflows involve software engineering, client-facing judgment, or coordination-intensive production, the 87% figure does not describe their situation. Agent productivity claims require a task-category filter before they translate into staffing or tooling decisions.
The study’s own caveats are worth keeping visible: early adopters skew toward AI-native users, the observation window covers only the first three months after Computer’s February 2026 launch, and the efficiency estimates depend on assumptions about human-equivalent tool time that carry meaningful uncertainty bands.
Before your procurement team uses these numbers to justify a budget decision, run a controlled sample of your actual task mix against the baseline. The study gives you the methodology to do exactly that.
Perplexity Research (research.perplexity.ai), 2026-06-08.
