Perplexity Computer Cuts Knowledge Work Time by 87% in Harvard Study

Perplexity and Harvard publish the first large-scale study showing AI agents cut task time by 87% and cost by 94% versus search-assisted humans

·
·
·
Perplexity Computer Cuts Knowledge Work Time by 87% in Harvard Study
  • 87% less time, 94% lower cost: Perplexity Computer completes matched tasks in 36 minutes vs. 269 minutes for Search + human execution.
  • Harvard collaboration: The paper uses 10,000 matched query pairs as natural experiments to compare the same tasks across both products.
  • Quality improves with autonomy: Dissatisfaction rates were 55% lower on Computer (1.3%) than Search (2.9%) on matched sessions.
  • Scope expansion is the key finding: 23% of Computer queries involve tasks that never appeared in the same users' Search history -- entirely new work.
  • Cross-domain work surges: Computer users work outside their primary occupation 59% of the time vs. 50% for Search; 51% of queries span 3+ knowledge domains.
  • Rapid adoption: Cumulative Computer queries grew 84x from week one; usage is led by research/analysis (25.8%) and document creation (18.6%).

Most AI productivity research has focused on chatbots helping people write better emails or debug code faster. A new study from Perplexity and Harvard Business School goes further, asking a different question: what happens when you stop giving humans better answers and start giving them an agent that does the work itself?

The answer, based on three months of real production data, is striking. Workers using Perplexity Computer -- a general-purpose agent that plans, browses, codes, writes, and connects to external services autonomously -- completed the same tasks in 87% less time at 94% lower cost compared to humans using Search alone. And the quality went up, not down.

The Gap Between Answering and Doing

The core tension in AI-assisted work has always been the handoff. A search engine or chatbot gives you the answer. You still have to open the right tools, pull the files, run the code, edit the document, and decide what to do next. Every step is yours to execute.

Agents change that division of labor. A user specifies an outcome, and the system plans across tools, executes intermediate steps, asks for missing inputs when needed, and returns a finished deliverable. The user moves from operator to supervisor. That shift is what this study tries to measure -- rigorously, at scale, in the wild.

Perplexity launched Computer in 2026 as a general-purpose agent orchestrator that autonomously works toward user-specified objectives across complex environments and lengthy time horizons. The study compares it directly against Perplexity Search, which represents the classic "answer engine" paradigm.

How They Built the Comparison

The methodology is the most interesting part of the paper. You can't just compare random Search and Computer sessions -- the tasks are too different. So the researchers built a controlled comparison using matched pairs.

Across 10,000 matched pairs, the team identified sessions where the same user asked near-identical queries to both Search and Computer, and where Computer ran at least one execution tool. This gives you a natural experiment: the same underlying task, attempted with two different tools.

To estimate how long a human would take to complete the manual steps that Computer handles automatically, they used three independent methods:

  • Tool-based estimate: Computer tool calls were classified into "Search" tools (information retrieval already handled by Search) and "Do" tools (execution steps a human would need to perform manually). Per-call time estimates approximate the time a skilled professional would spend performing the equivalent action manually.
  • LLM-based estimate: Queries from Computer sessions were fed into an LLM to estimate the time required for a skilled professional who receives answers from Search but must perform all execution steps manually.
  • User interviews: 25 semi-structured interviews with active Computer users across domains elicited pre-Computer workflows and the time those workflows would have taken.

Human labor costs were priced using domain-specific average hourly wages from U.S. Bureau of Labor Statistics Occupational Employment and Wage Statistics May 2025 data. All three methods pointed in the same direction.

The Numbers Are Hard to Ignore

Under the tool-based estimate, the average Search + Human task takes 269 minutes, while the corresponding Computer + Human workflow takes 36 minutes -- an 87% reduction in task time. Translating that to cost, Computer reduces estimated task cost by 94% on average.

Bar chart comparing task time and cost between Search+Human and Computer+Human across categories

The efficiency advantage appears across all 18 domains, with 79-92% time savings and 87-96% cost savings. Programming is the most extreme case: 596 minutes for Search + Human versus 48 minutes for Computer + Human -- a 92% time reduction that yields a 96% cost reduction.

The LLM-based estimate yielded similar results: 84% time reduction and 93% cost reduction. User interviews showed a wider range -- from 5x to over 300x speedups -- with a median of 25x. The consistency across three independent estimation methods is what makes the finding credible.

More Autonomy, Better Quality

The natural worry with autonomous agents is that more automation means more errors. The data says otherwise. On matched multi-turn sessions, meaningful next-turn dissatisfaction was 1.3% for Computer versus 2.9% for Search -- a 55% reduction. Any dissatisfaction, including mild signals, was 10.8% for Computer versus 16.6% for Search.

The machine work gap is enormous. Computer performs 26 minutes of machine execution per session on average, versus 33 seconds for Search -- a 48x increase in machine work on effectively the same tasks. Yet abandonment rates were nearly identical: 3.7% of Computer sessions and 3.4% of Search sessions contained at least one user stop event.

Computer also reaches much further into a user's connected services. 7.9% of Computer sessions make at least one connector call (through MCP or API endpoints) versus 1.8% of Search sessions, and Computer averages 1.19 connector calls per session versus 0.10 for Search -- a 12x ratio. In other words, Computer doesn't just run longer; it pulls data and takes actions across more of a user's connected apps.

The Scope Shift Is the Real Story

Speed and cost savings on matched tasks are impressive, but the most consequential finding is about what kinds of work people attempt in the first place. Agents don't just make existing tasks faster -- they make previously impractical tasks viable.

Computer queries are more cognitively complex than Search queries. On Bloom's Revised Taxonomy (a framework for classifying cognitive tasks from basic recall to creative synthesis), 76% of Computer queries require higher-order cognition, compared to 55% for Search. 50% of Computer queries are Create-level tasks, compared to 26% for Search.

Bar chart comparing knowledge domains per query between Search and Computer

Computer tasks draw on more knowledge domains. The average Computer task requires substantive expertise in 2.40 O*NET Knowledge areas, compared to 1.74 for Search -- a 38% increase. Computer is nearly three times as likely as Search to require three or more knowledge domains: 51% versus 17%.

The cross-occupation data is equally striking. Across a sample of 8,000 users from eight occupation clusters, Computer users work outside their primary occupation 59% of the time, compared to 50% for Search. A finance professional using Computer is regularly doing work that would previously require a developer, a designer, or a legal analyst.

Perhaps the most telling metric: 23% of Computer queries engage at least one task activity that never appeared in the same users' Search history. These are tasks users simply weren't attempting before. They concentrate in software and web development, documentation production, and data visualization -- exactly where autonomous execution matters most.

Growth and Adoption

Cumulative Computer queries reached 84x their first-week total by the end of the study window. As a baseline, cumulative Search usage among Computer users reached 14x its first-week total, higher than the 12x growth for non-Computer users. Agents aren't replacing search -- they're expanding total usage.

Line chart showing 84x cumulative growth in Computer queries versus 14x for Search among Computer users

In a random sample of 100,000 Computer queries, Research and Analysis was the largest task category at 25.8%, followed by Document and Asset Creation at 18.6%. Use was distributed across software, finance, marketing, business operations, healthcare, education, legal, and media.

Where It Falls Short

The paper is honest about its limitations. The observation window covers early adopters who skew toward AI-native users, so patterns may shift as adoption broadens. The efficiency estimates depend on assumptions about how long manual steps take -- though sensitivity analyses show the advantage holds even under conservative assumptions. And the study only captures behavior within the Perplexity ecosystem, missing whatever else users do outside it.

Computer is not a replacement for specialized tools -- it's a powerful generalist. If you need deep expertise in a specific domain, dedicated tools will likely outperform it. The value proposition also requires regular, substantial use: the Computer agent requires a Max subscription at $200/month.

What This Actually Means

Prior research on AI at work -- from the Noy and Zhang writing study to the BCG consulting experiments -- focused on AI as a co-pilot: a tool that makes individual steps better. This study is measuring something different: what happens when the agent handles the steps entirely, and the human just sets the goal and reviews the result.

The finding that 23% of Computer tasks were entirely new -- not just faster versions of existing work -- is the number that matters most for thinking about the future of knowledge work. It suggests the economic impact of agents won't just show up as productivity metrics on existing jobs. It will show up in how work gets bundled, what roles exist, and what a single person can accomplish without a team. The full technical report is available on arXiv.

Trending
  • No trending articles

Comments

avatar

Next Reads