Perplexity says users of Computer, its general-purpose AI agent, are delegating longer, tool-heavy workflows and attempting more complex work than users of Search.

For AI program owners, product leaders, managers, developers, analysts, and operations teams, the decision is whether agents should be treated as faster task tools or as systems that change which work is worth taking on.

One reason is the difference in execution. In matched tasks, Computer performed 26 minutes of machine work per session on average, compared with 33 seconds for Search.

In short, Perplexity's study says agents move users from doing each step to defining goals, adding context, checking output, and asking for extensions. The productivity gains are real, but the stronger finding is about ambition. People appear to attempt more complex and cross-domain work when an agent can carry the execution load.

An AI agent is a system that can plan across tools, perform intermediate steps, ask for needed input, and return a completed deliverable instead of only answering a prompt.

AI agents can take a user goal, plan the work across tools, and return a finished output that a person reviews.

Perplexity and Harvard Business School researchers studied how people use Perplexity Computer, a general-purpose agent orchestrator, alongside Perplexity Search. Search is the company's answer engine for asking questions and receiving cited answers. Computer works differently. It is an agent designed to pursue user-specified objectives across complex environments and longer time horizons.

The study compares the products by autonomy, efficiency, and the kinds of work users attempted. Autonomy measures how much work the system performs without human intervention. Efficiency measures estimated time and labor-cost savings relative to Search. The central question is whether Computer changes what users choose to take on.

Computer launched on February 25, 2026, and Perplexity says usage grew quickly during its first three months. By May 27, cumulative Computer queries reached 84 times their first-week total. During the same period, cumulative Search usage among Computer users reached 14 times its first-week total, compared with 12 times for non-Computer users. Perplexity also reports that, after matching users by subscription tier, primary Search topic, and prior Search intensity, Computer adopters made 1.05 more Search queries per day than similar non-Computer users.

The reported use cases also lean toward finished work. 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%. Perplexity says the task mix skewed toward generative work such as documents, spreadsheets, codebases, websites, and workflows that require multiple tools.

The adoption data indicates that Computer added another work mode rather than replacing Search. Users still relied on Search for cited answers, while Computer handled longer work that required planning, tool use, and completed outputs.

Perplexity's most direct measure of agent autonomy is machine execution time. In 10,000 matched pairs of near-identical initial queries, Computer performed 26 minutes of machine execution per session on average, compared with 33 seconds for Search. Perplexity calls that a 48x increase in machine work on effectively the same tasks. Computer also ran longer in the median comparison, at 9 minutes for Computer and 14 seconds for Search, or 40 times longer.

The longer run time did not come with a large increase in user stops. Perplexity reports that 3.7% of Computer sessions and 3.4% of Search sessions had at least one user stop event. Computer did pause for user input more often. The source says 13% of Computer queries invoked at least one pause-for-user tool, compared with 0.3% for Search. Those pauses usually asked for approval or clarification.

Computer also used more connected tools. Perplexity reports that 7.9% of Computer sessions made at least one connector call through Model Context Protocol or API endpoints, compared with 1.8% of Search sessions. Computer averaged 1.19 connector calls per session, compared with 0.10 for Search.

Computer takes a desired outcome and runs through more of the intermediate steps itself. It can search, browse, write, edit, run code, check intermediate results, and call connected services. The user still has work to do, but that work moves toward setting the goal, granting permission, reviewing the result, and asking for revisions or extensions.

Follow-up behavior fits that pattern. In a 1,000-pair multi-turn sample, Perplexity reports almost identical rates of follow-ups that moved the task forward, at 52.7% for Computer and 52.9% for Search. Computer users asked for more extensions and spent slightly more follow-up activity reviewing and revising output, while Search users used more short directives such as confirmations, retries, and format requests.

Perplexity also reports lower dissatisfaction in matched multi-turn sessions. Meaningful dissatisfaction in the next user turn was 1.3% for Computer and 2.9% for Search, which the source describes as a 55% reduction. Any dissatisfaction, including mild indicators, was 10.8% for Computer and 16.6% for Search.

These efficiency figures measure something different from the machine-execution times above. The 26-minute and 33-second numbers are how long the system itself ran. The figures below estimate total human-equivalent task time: how long the whole task would take a person, including the manual work a human would otherwise do.

Perplexity estimates efficiency by comparing Search plus human manual execution with Computer plus human goal-setting and review. The company says it cannot directly observe how long each task would take a human, so it triangulates with 3 methods. It uses a tool-based estimate of manual execution time, an LLM-based estimate, and interviews with 25 active Computer users.

Under the tool-based estimate, the average Search plus human task took 269 minutes, while the corresponding Computer plus human workflow took 36 minutes. Perplexity reports that as an 87% reduction in task time. Combining model cost with domain-specific human labor cost, the study estimates a 94% average reduction in task cost.

The largest reported programming case was 596 minutes for Search plus human work, compared with 48 minutes for Computer plus human review. Perplexity reports that as a 92% time reduction and a 96% cost reduction. Across all 18 domains, Perplexity reports 79% to 92% time savings and 87% to 96% cost savings.

The efficiency numbers make the story larger than speed alone because they show how cheaper execution changes the cost of trying.If a task that once required hours of manual work can be defined and reviewed in a shorter cycle, more projects become reasonable to attempt. The relevant business question becomes which workflows deserve the agent's effort and who is qualified to judge the result.

The strongest evidence for changed work behavior comes from what users attempted with Computer. Perplexity reports that Computer users worked outside their main professional area 59% of the time, compared with 50% for Search. The largest increases appeared in Management and Entrepreneurship, Digital Technology, Arts and Design, and Healthcare and Human Services.

The study also says Computer queries required more complex thinking. In a sample of 5,000 Computer queries and 5,000 Search queries from the same dual-product users, 76% of Computer queries involved higher-level thinking such as analysis, evaluation, or creation, compared with 55% for Search. Half of Computer queries were Create-level tasks, compared with 26% for Search.

Computer tasks also drew on more knowledge areas. Perplexity used O*NET, a U.S. Department of Labor occupational database, to measure how many knowledge categories each task required. The average Computer task required 2.40 knowledge areas, compared with 1.74 for Search, a 38% increase. Computer was also nearly 3 times as likely as Search to require 3 or more knowledge domains, at 51% versus 17%.

One of the clearest measures is the set of tasks that appeared in Computer but not in Search among the same users. Using Perplexity's most conservative measure, 23% of Computer queries engaged at least one task statement that never appeared in the paired Search sample. Those Computer-only activities concentrated in software and web development, documentation production, and data visualization or graphics.

For businesses, that is the heart of the story. Agents may make work faster, but the reported behavior also suggests that people use agents to cross into work they previously would have avoided, delayed, or handed to someone else.

Perplexity is reporting on its own product, and it names several limits. The observation window is early, and early adopters skew toward AI-native users. Users were also experimenting and changing workflows while the product was evolving.

The matched-query design has another limit. It compares Search and Computer sessions that start with near-identical queries, but Perplexity says this leaves out many Computer tasks with no close Search equivalent. Sessions are also an imperfect way to define task units because users do not always work in clean session boundaries.

The efficiency figures require the most care. Perplexity's estimates depend on assumptions about how long a human would take to perform tool-equivalent work and how much time a person spends overseeing Computer. Perplexity says LLM-based estimates and user interviews point in the same direction, and sensitivity checks keep Computer ahead, but the exact numbers should be read as approximate.

The study also observes behavior only inside the Perplexity ecosystem. It does not capture what users did in other tools or how work changed outside Perplexity's products.

Perplexity's study puts agent adoption in the work-design layer of AI use. The reported gains cover more than speed on the same task. They show what happens when a system can carry more of the workflow and the user can spend more time deciding what outcome is worth pursuing.

That is a different operating question for organizations. A manager using agents may ask which projects become viable with a smaller team. A developer may decide which build, research, or documentation work can be delegated safely. A marketer or analyst may test more campaign, research, or reporting ideas because the manual assembly work takes less time.

The study does not predict how labor markets or team structures will change. Perplexity itself says the long-run impact will show up in how work is bundled, how roles are defined, and how teams are structured. For now, the practical finding is closer to the desk level: people appear to take on wider and more complex work when an agent handles more of the execution.

Q: What is Perplexity Computer?
A: Perplexity Computer is a general-purpose AI agent designed to pursue user-specified objectives across complex environments and longer time horizons.

Q: How is Perplexity Computer different from Search?
A: Search answers questions and helps users synthesize information. Computer performs more of the workflow, including searching, browsing, writing, editing, running code, checking results, and calling connected services.

Q: What did Perplexity find about AI agents and knowledge work?
A: Perplexity says Computer users delegated longer workflows, relied on more machine execution, and attempted more complex and cross-domain work than Search users on comparable tasks.

Q: Do AI agents make people attempt harder work?
A: In Perplexity's data, Computer queries more often involved complex thinking, Create-level tasks, multiple knowledge areas, and work outside the user's main professional area.

Q: Are AI agents mainly productivity tools?
A: Productivity is part of the finding. Perplexity reports large estimated time and cost reductions, but the larger business question is whether lower execution effort changes what people and organizations choose to attempt.

Q: What should companies consider before adopting AI agents?
A: Companies should decide where agents can extend what a person or team can take on, and where the work still requires human execution or expert review from the start.

Q: What are the limits of Perplexity's study?
A: The study is based on Perplexity usage, covers an early observation window, includes early adopters, and relies on estimates for human-equivalent work time. Perplexity says the exact efficiency numbers should be treated as approximate.

Perplexity's study shows agent adoption is becoming a work-design decision. When an agent can search, browse, code, edit files, connect to services, and produce deliverables, teams have to decide which work becomes worth attempting.

Lower execution burden can make research, analysis, documentation, software, and business workflows more practical for individuals and teams that previously lacked the time, staff, or specialist help to attempt them.

As agents handle more execution, AI program owners, operations leaders, developers, analysts, marketers, educators, and managers may need to apply more human judgment to choosing the right goal, supplying the right context, checking the agent’s output, and deciding when that output is good enough for use.

Agent tools are moving from answering questions to performing work across tools, so adoption decisions should include workflow design, review habits, permissioning, and quality checks alongside speed and cost expectations.

Organizations need to decide where AI should extend what a person or team can take on, and where the work still needs human execution or expert review from the start.

In short, the business case for agents may depend on the new work they make practical as much as the minutes they save. If execution becomes cheaper, the value of knowledge work moves closer to goal selection, judgment, and review.

As AI agents take on more of the workflow, speed becomes only part of the story. The longer-term question is which work people and teams now believe is possible.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing support, AEO/GEO/SEO optimization, image concept development, and editorial structuring support from ChatGPT, an AI assistant. All final editorial decisions, perspectives, and publishing choices were made by Alicia Shapiro.