How AI Agents Reshape Knowledge Work: Autonomy, Efficiency, and Scope
Using a task-based framework, real-world data from Perplexity shows AI agents significantly boost automation, efficiency, and task scope, with productivity gains of up to 87%.
Key Findings
Methodology
This study employs an individual-level task framework, integrating real production data from Perplexity’s Search and Computer products. It models tasks as sequences of atomic steps with increasing value and costs associated with different modes—conversational and autonomous agent. The analysis involves natural experiments by matching sessions with nearly identical initial queries, controlling for user and task heterogeneity. It compares autonomous task planning and execution, classifies query complexity and scope, and applies a cost-value optimization model based on a knapsack formulation to predict how agent access expands feasible tasks and enhances total value. The methodology combines empirical matching, classification, and theoretical modeling to quantify the impact of autonomous AI agents on work productivity and scope.
Key Results
- In matched sessions, Computer autonomously performs 26 minutes of task planning and execution per session, compared to 33 seconds for Search, representing a 48× increase in machine work. This automation shifts follow-up queries toward higher-order activities like verification and extension, with dissatisfaction rates 55% lower than Search, indicating improved quality.
- For the same tasks, Computer reduces total completion time from 269 minutes to 36 minutes, achieving an efficiency gain of 87% and lowering estimated costs by 94%. Sensitivity analyses confirm the robustness of these findings across different human time estimates and independent validation methods.
- Computer significantly broadens the scope of work: it crosses occupational boundaries more frequently (by 9 percentage points), involves higher cognitive complexity (71% non-routine tasks vs. 53%), draws on a wider set of expertise (average 2.4 vs. 1.7 knowledge domains), and combines multiple subtasks into single queries, enabling new work activities previously absent in Search usage.
Significance
This research provides the first large-scale empirical evidence of AI agents’ economic impact in real-world knowledge work. It demonstrates that autonomous agents can dramatically accelerate workflows, improve output quality, reduce costs, and expand work scope. These findings address longstanding questions about AI’s productivity effects, offering insights into how automation reshapes occupational tasks, organizational structures, and economic value. The results have profound implications for industry adoption, policy formulation, and future AI development, emphasizing the transformative potential of autonomous AI systems in diverse professional contexts.
Technical Contribution
The paper introduces a task-step based cost-value model that captures the fixed costs of delegation and the marginal benefits of automation, providing a theoretical foundation for understanding task boundary expansion. It combines empirical matching with classification and modeling to validate the predictions that autonomous agents extend feasible task sets, increase total value, and improve value-to-cost ratios. The approach bridges economic theory with large-scale real-world data, offering a novel framework for analyzing AI’s productivity effects and work reorganization at the individual task level.
Novelty
This study is pioneering in empirically quantifying the economic and scope effects of autonomous AI agents in real-world settings. Unlike prior work limited to controlled experiments or theoretical models, it leverages large-scale production data to demonstrate how autonomous agents expand the feasible task frontier, enhance productivity, and enable new forms of work. Its integration of a task-based economic model with actual user data provides a comprehensive understanding of AI-driven work transformation, marking a significant advance over existing literature.
Limitations
- The analysis relies on data from Perplexity’s specific products, and generalizability to other platforms or industries requires further validation;
- Task value is not directly measured, and inferences are based on cost and scope proxies, which may introduce bias;
- The cost model assumes fixed and marginal costs with simplified relationships, not accounting for learning effects or dynamic system improvements; future research should incorporate broader datasets and more nuanced cost structures.
Future Work
Future research should extend to multiple industries and task types, validating the universality of these findings. Incorporating direct measures of task value and user satisfaction will refine the economic impact assessments. Developing multi-modal, multi-layered autonomous systems can further enhance scope and efficiency. Additionally, exploring ethical, fairness, and control mechanisms for autonomous agents will be crucial to ensure sustainable and responsible deployment. Long-term studies on organizational and occupational restructuring driven by AI agents are also needed to understand broader societal implications.
AI Executive Summary
The landscape of knowledge work is undergoing a profound transformation driven by advances in artificial intelligence. Historically, tasks such as research, analysis, and content creation required significant human effort, time, and expertise. Traditional AI tools, like search engines and simple assistants, provided support but lacked autonomy and integration, limiting their productivity impact. Recent breakthroughs in generative models and multi-step reasoning have paved the way for autonomous AI agents capable of executing complex workflows end-to-end.
This paper investigates the real-world impact of such agents through data from Perplexity’s Search and Computer products. Search, introduced in 2022, exemplifies a conversational assistant that provides answers based on knowledge synthesis. In contrast, Computer, launched in 2026, is a general-purpose autonomous agent that can plan, browse, code, create documents, and delegate subtasks without human intervention. The core question is: how does this shift from interactive assistance to autonomous execution influence productivity, work scope, and costs?
The authors develop a task-based economic framework, modeling tasks as sequences of steps with increasing value and costs associated with different modes—conversational and autonomous. This model predicts that autonomous agents expand the feasible task set, enabling users to undertake more complex, higher-value work. Empirical analysis using matched session data reveals that Computer performs 26 minutes of autonomous work per session, compared to just 33 seconds for Search, a 48-fold increase. This automation reduces total task completion time from 269 minutes to 36 minutes, improving efficiency by 87% and cutting costs by 94%. Moreover, Computer broadens the scope of work: users engage in more cross-occupational, cognitively demanding, and multi-faceted tasks, often combining multiple subtasks into single queries.
These findings demonstrate that AI agents not only accelerate workflows but also enhance output quality and expand work boundaries. They facilitate the automation of complex, knowledge-intensive activities, making previously costly or infeasible tasks accessible. This has significant implications for industries seeking to boost productivity and innovate work processes. The study also highlights the economic advantage of the cost structure—higher fixed costs balanced by lower marginal costs—enabling the scaling of high-value tasks.
Despite these promising results, limitations remain. The analysis is based on data from a specific platform, and direct measurement of task value remains challenging. Future research should validate these findings across diverse settings, incorporate direct value assessments, and explore multi-modal, multi-layered autonomous systems. Addressing ethical and control issues will be vital for responsible deployment. Overall, this research provides compelling evidence that autonomous AI agents are poised to reshape the future of knowledge work, driving efficiency, quality, and scope to new heights.
Deep Analysis
Background
The evolution of AI has transitioned from simple rule-based systems to sophisticated generative models like GPT-4, which have demonstrated remarkable capabilities in natural language understanding, reasoning, and content creation. Early AI applications in knowledge work primarily involved information retrieval and basic automation, exemplified by search engines and rule-based assistants. The advent of large language models (LLMs) has enabled more complex interactions, but these systems largely remained interactive, requiring human oversight and iterative prompting. Recent developments have introduced multi-step reasoning, tool invocation, and autonomous planning, exemplified by systems like Schick et al. (2023) on tool use, Kwa et al. (2025) on time horizon metrics, and Sarkar (2026) on code automation. Despite these advances, empirical evidence on how autonomous agents impact real-world productivity, task scope, and organizational structures remains limited. Most prior work focused on controlled experiments or theoretical models, leaving a gap in understanding their practical economic effects at scale. This study bridges this gap by analyzing large-scale production data from Perplexity, providing insights into how AI agents reshape work in diverse contexts.
Core Problem
While AI models have demonstrated impressive capabilities in isolated tasks, integrating these into real-world workflows poses significant challenges. Key issues include quantifying productivity gains, understanding how autonomous agents influence task complexity and scope, and assessing economic impacts such as cost reduction and value creation. Existing literature offers limited empirical evidence, especially regarding long-term effects and work reorganization. The core problem is to determine whether autonomous AI agents can reliably perform complex, multi-step tasks at scale, and how their deployment alters occupational boundaries, skill requirements, and organizational processes. Addressing these questions is crucial for guiding industry adoption, policy development, and future AI research aimed at maximizing societal benefits while mitigating risks.
Innovation
This research introduces several innovations: 1) a task-step based economic model that captures fixed and marginal costs, predicting how autonomous agents expand feasible task sets; 2) a large-scale empirical comparison of Search and Computer, controlling for user and task heterogeneity through session matching; 3) a detailed classification of query complexity and scope, revealing how autonomous agents enable higher-order, cross-occupational, and multi-subtask work. Unlike prior studies limited to controlled experiments or theoretical analyses, this work leverages real-world data to validate the model’s predictions, providing a comprehensive understanding of AI’s productivity and work reorganization effects. The integration of economic modeling with large-scale behavioral data marks a significant methodological advance, offering a blueprint for future research in AI-driven work transformation.
Methodology
- �� Task modeling: Tasks are represented as sequences of steps (s_j), with value (v_j) increasing with step count, assuming indivisibility and full realization upon completion;• Cost structure: Fixed costs differ between modes—f_conversational for simple queries, f_agent for autonomous tasks—while marginal costs per step (m_conversational, m_agent) reflect the efficiency advantage of automation;• Empirical matching: Sessions with similar initial queries are identified via cosine similarity (>0.99), forming pairs for comparison;• Classification: Queries are categorized by complexity, scope, and occupational boundaries, using large-scale labeling and expert annotation;• Theoretical derivation: A knapsack-based model predicts how agent access expands feasible tasks, increases total value, and improves value-to-cost ratios, with formal proofs provided.
Experiments
- �� Data sources: 2026年2月至5月的生产数据，涵盖7个样本，包括用户查询、会话匹配、任务分类和行为分析；• 核心设计：匹配相似会话对，控制用户异质性，比较Computer与Search在相似查询上的表现；• 评估指标：自主执行时间、用户满意度、任务范围、认知复杂度、知识领域覆盖率；• 分类方法：采用余弦相似度筛选高相似会话，结合专家标注进行任务分类，分析认知复杂度和跨界行为；• 统计分析：差异检验、敏感性分析、多模型验证，确保结论稳健。
Results
- �� Computer在匹配会话中自主执行任务时间达26分钟，而Search仅33秒，效率提升近48倍；• 任务完成总时间由269分钟降至36分钟，效率提升87%，成本降低94%；• Computer显著扩展工作范围，跨越职业界限比例高出Search 9个百分点，认知复杂度和知识领域显著增加，且能将多个子任务合成单一查询，开启新型工作活动。
Applications
- �� 立即应用：企业可利用AI代理优化客服、内容生成、数据分析等流程，降低人力成本，提高响应速度；• 长远愿景：推动智能自动化在科研、法律、医疗等领域深度融合，实现长周期、复杂任务的自主完成，重塑职业结构和工作生态。
Limitations & Outlook
- �� 目前研究依赖特定平台数据，跨行业验证尚需展开；• 任务价值未直接测量，主要依赖成本和范围指标推断，存在偏差；• 模型假设简化了成本关系，未充分考虑学习效应和系统优化带来的动态变化。未来应结合多源数据和多场景验证，完善理论模型，探索更复杂的成本价值关系。
Plain Language Accessible to non-experts
Imagine you’re in a kitchen preparing a big meal. In the past, you had to do everything yourself—shopping, chopping, cooking, plating—step by step, which took a lot of time and effort. Now, if you had a super-smart chef robot that could automatically chop vegetables, cook dishes, and even plan the menu, all you need to do is tell it what you want to eat. It handles most of the work, saving you hours and letting you focus on enjoying the meal.
This robot is like an AI assistant that can understand your needs and plan out each step, then execute it without your constant input. You just give it a goal, and it autonomously manages the entire process. This means you can make more complex dishes, try new recipes, or serve a larger crowd—all with less effort.
In the workplace, it’s similar. Traditionally, people spent hours doing repetitive or complex tasks—research, analysis, report writing. Now, with AI agents like Perplexity’s Computer, these tasks can be automated. The AI plans, browses, codes, and creates documents on its own, freeing workers to focus on higher-level thinking and decision-making. This shift makes work faster, cheaper, and opens up new possibilities for what can be achieved, just like having a robotic chef in your kitchen transforms cooking into a more creative and efficient activity.
ELI14 Explained like you're 14
Imagine you’re at school doing homework. Usually, you spend a lot of time looking up facts, solving problems, and writing answers. Sometimes, you get stuck or it takes forever. Now, think about having a super-smart friend who can help you instantly. You tell him what the question is, and he automatically finds the answer, explains it, and even helps you organize your notes. You still learn, but it’s much faster and easier.
This friend is like a clever robot that understands your questions and can do the work for you. Instead of spending hours on each problem, you just ask, and it handles everything. This way, you can finish your homework quickly and have more time to play or learn new things.
In jobs, it’s similar. People used to spend hours searching for information, writing reports, or analyzing data. Now, with AI helpers, these tasks can be done automatically. The AI can plan, browse the internet, write code, and prepare documents without much human help. It makes work faster, cheaper, and allows us to do more complex and interesting projects. As these AI helpers get smarter, they’ll help us learn and work in ways we never imagined, making everything more fun and productive!
Abstract
Frontier AI systems are bridging the gap between intelligence and utility by shifting from conversational assistants to autonomous agents that execute tasks end to end. Using production data from Perplexity's Search and Computer products, we study this transition by examining how AI agents accelerate and reshape knowledge work. Three key empirical findings emerge. First, using sessions with near-identical initial query pairs as natural experiments for the same underlying task attempted with both products, Computer performs 26 minutes of autonomous work per user session, versus 33 seconds for Search. Computer automates task decomposition and execution that Search users might otherwise manually orchestrate and implement. As a result, Computer shifts follow-up query distribution toward higher-order work such as verification and extension. Autonomy also increases execution quality, with per-query dissatisfaction rates 55% lower on Computer than on Search. Second, due to its autonomy advantage, Computer reduces completion time from 269 to 36 minutes on matched tasks, lowering estimated time and cost by 87% and 94%, respectively, compared to humans equipped with Search alone. Third, Computer changes the scope of work that users attempt: Computer queries more often cross occupational boundaries, require higher-order cognition, draw on broader expertise, take the form of composite tasks that bundle interdependent subtasks into a single query, and unlock work activities that are essentially absent from Search usage among the same users. Together, the evidence indicates that AI agents accelerate workflows, enhance output quality, reduce costs, and expand the breadth and depth of automated work.