How AI Agents Reshape Knowledge Work: Autonomy, Efficiency, and Scope
This paper investigates how the transition from conversational AI assistants to autonomous agents changes the way people perform knowledge work. By analyzing production data from Perplexity’s Search and Computer products, the authors examine how shifting from interactive, step-by-step guidance to end-to-end autonomous execution impacts productivity, task quality, and the types of work users are willing to undertake.
The Shift to Autonomous Agents
The research distinguishes between conversational assistants, which support isolated information exchange, and autonomous agents, which can plan and execute tasks across multiple tools with minimal human intervention. The authors propose a framework where agents have higher fixed costs—due to the need for clear objective setting and review—but significantly lower marginal costs for each step of a task. This structure suggests that while conversational tools remain efficient for quick, simple queries, agents become the preferred choice for complex, multi-step projects where they can automate the heavy lifting of execution.
Efficiency and Quality Gains
The study uses matched sessions—where users performed nearly identical tasks using both Search and Computer—as natural experiments. The results show that Computer performs 26 minutes of autonomous work per session, compared to just 33 seconds for Search. This automation leads to substantial efficiency gains: on matched tasks, the Computer-assisted workflow reduces the average completion time from 269 minutes to 36 minutes. Furthermore, the autonomy provided by the agent does not come at the cost of quality; the dissatisfaction rate for queries handled by the Computer agent was 55% lower than that of the Search product.
Expanding the Scope of Work
Beyond simply speeding up existing tasks, the use of autonomous agents fundamentally changes the nature of the work users attempt. The data shows that users of the Computer agent are more likely to cross occupational boundaries, moving into domains outside their primary expertise. These queries are also more cognitively complex, frequently requiring the synthesis of multiple knowledge domains and the bundling of interdependent subtasks into a single request. By automating the generative and executional components of these tasks, agents allow users to tackle higher-order work that was previously absent from their workflows.
Economic Implications
The findings suggest that autonomous agents act as a force multiplier for knowledge workers. By reducing the time and cost required to complete complex, multi-step tasks, agents lower the barrier to entry for specialized work. This shift not only improves individual productivity but also suggests broader changes for organizational structures, as workers become capable of absorbing tasks that previously required coordination across different roles or departments. The authors note that while these tools significantly expand the breadth and depth of automated work, the ultimate impact depends on the balance between the agent's ability to execute and the human's ability to verify the final output.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!