AI Agents Are Reshaping the Scope of Knowledge Work, Study Finds

Researchers working with Perplexity AI and Harvard Business Review have published a study showing that autonomous AI agents don't just speed up office work, they change its very nature. A comparison of Perplexity's classic conversational assistant, Search, with the autonomous agent Perplexity Computer found that agent users more often take on tasks outside their own professional field and engage in more complex, multi-step projects.

From Operator to Supervisor

The key difference between a classic chatbot and an agent lies in what reaches the user at the end of the process. A conversational assistant delivers information that a person still has to process and turn into a finished output. An autonomous agent plans the steps, uses multiple tools on its own, and returns a finished product that only needs verification.

Researchers on the Perplexity team - Jeremy Yang, Kate Zyskowski, Noah Yonack and Jerry Ma - describe this as a shift in the user's role from operational executor toward supervisor of autonomous systems. Instead of manually carrying out each step of a task, people increasingly evaluate and refine the finished output produced by the agent.

Agents are shifting AI use from searching for and synthesizing information toward planning and independently carrying out tasks - from Perplexity's research report

How the Efficiency Was Measured

The methodology relied on natural experiments: researchers paired sessions with nearly identical initial queries but different tools, some going to Search, others to Computer. This let them compare completion time, cost and quality on matched tasks without needing an artificial lab experiment.

On matched tasks, completion time dropped from 269 to 36 minutes when moving from the Search-plus-human model to the Computer-plus-human model. The user dissatisfaction rate with results fell by 55 percent, from 2.9 percent for Search to 1.3 percent for Computer. Adoption grew quickly - the volume of Computer queries reached 84 times its first-week level by May 27, 2026.

A Broader Range of Skills

The study also measured the cognitive complexity of tasks using the O*NET classification, the standard US system for describing occupations and required competencies. Tasks assigned to the Computer agent required knowledge from an average of 2.40 fields versus 1.74 for Search, a 38 percent increase. This means users turn to the agent for tasks that combine several areas of knowledge at once, rather than sticking to a narrow specialization.

The authors stress that this shift has organizational consequences. Since the agent takes on both breaking down the task and carrying it out, employees can focus on higher-order activities, verifying results and expanding the scope of a project, instead of laboriously executing each step themselves.

What It Means for Companies

For HR departments and executives, this means rethinking roles, workflows, accountability and oversight of output generated by agents. The Harvard Business Review article notes that leaders need to build quality-control mechanisms for agent work if they want the growing autonomy of AI systems to translate into real business value rather than a loss of control over processes.

The study fits into a broader trend seen at technology companies, where coding and research agents are starting to take on entire multi-step tasks instead of single commands. Perplexity, known primarily for its AI-powered search engine, is developing its Computer product in this direction as a tool for independently carrying out office tasks, competing with similar initiatives from major AI labs.

For Polish companies considering deploying similar tools, the study offers concrete numbers for calculating return on investment - cutting task completion time by nearly 90 percent and reducing costs by more than 90 percent are arguments that translate easily into a department budget. At the same time, the lower dissatisfaction rate suggests that agent work quality doesn't have to come at the expense of speed.