This research paper from Harvard Business School and Perplexity AI provides the first field evidence on how autonomous AI agents reshape knowledge work compared to conversational AI assistants. Using production data from Perplexity's Search and Computer products, the authors develop a task-based economic framework and examine four dimensions: adoption, autonomy, efficiency, and scope. The key finding is that Computer performs 26 minutes of autonomous work per session versus just 33 seconds for Search — a 48× increase — while reducing per-query dissatisfaction rates by 55%. On matched tasks, the agent reduces completion time from 269 to 36 minutes, yielding 87% time savings and 94% cost savings compared to humans using Search alone.
Beyond raw efficiency, the paper documents a fundamental expansion in the scope of work users attempt when given access to agents. Computer queries more frequently cross occupational boundaries, require higher-order cognition (71% abstract non-routine tasks vs. 53% for Search), draw on broader expertise domains (+38% more O*NET Knowledge domains per query), and bundle more subtasks into single queries. Notably, 23% of Computer queries involve task types that never appear in the same users' Search history, suggesting agents unlock entirely new categories of work rather than merely accelerating existing ones.
The conceptual framework elegantly captures the core tradeoff: agents have higher fixed delegation costs (specifying objectives, reviewing output) but lower marginal per-step execution costs, making them economically preferred for complex, multi-step tasks. User interviews corroborate the quantitative findings, with reported speedups ranging from 5× to 300× and cost reductions of two to three orders of magnitude. The paper also finds that Computer and Search are complements rather than substitutes — adopting Computer increases daily Search usage by approximately one additional query per day — consistent with the framework's prediction that agent-driven cost savings free up budget for shorter tasks where conversational tools retain an advantage.
A June 2026 paper from Perplexity researchers presents large-scale empirical evidence that autonomous AI agents fundamentally reshape knowledge work compared to conversational assistants. Using production data from Perplexity's Search (assistant) and Computer (agent) products, the study finds that agents perform 26 minutes of autonomous work per session versus just 33 seconds for assistants — a 48x increase. This autonomy translates into an 87% reduction in task completion time and a 94% reduction in cost, while simultaneously improving output quality, with per-query dissatisfaction rates 55% lower for the agent product.
Beyond speed and cost, the paper reveals that agents expand the scope of work users attempt. Computer queries more frequently cross occupational boundaries (59% vs. 50%), require higher-order cognition concentrated at the "Create" level of Bloom's taxonomy, and bundle multiple interdependent subtasks into single requests. Notably, 23% of agent queries engaged fine-grained work activities that users never attempted with the assistant, suggesting agents unlock entirely new categories of work. The authors frame these findings through a knapsack optimization model where agents' higher fixed delegation costs but lower marginal execution costs make them economically superior for complex, multi-step tasks.
The study's implications point toward a significant restructuring of knowledge work: humans shift from manual execution toward supervisory, verification, and extension roles, while agents handle planning and multi-step execution. The authors acknowledge important limitations including self-selection bias, a short observation window, and reliance on a single product ecosystem, but the magnitude of observed effects — validated through sensitivity analyses — suggests the transition from AI assistants to autonomous agents represents a structural shift in how knowledge work is organized and performed.
A new NBER working paper by researchers from OpenAI, Duke University, and Harvard provides the first comprehensive analysis of how people actually use ChatGPT, drawing on data from its launch in November 2022 through July 2025. By that point, ChatGPT had been adopted by roughly 10% of the world's adult population. The study finds that early adopters were disproportionately male, but the gender gap has since narrowed dramatically, and growth rates have been higher in lower-income countries.
Using a privacy-preserving automated classification pipeline on a representative sample of conversations, the researchers find that non-work usage has grown from 53% to over 70% of all conversations, outpacing the steady growth in work-related messages. Work usage skews toward educated users in highly-paid professional occupations. The three dominant conversation topics — "Practical Guidance," "Seeking Information," and "Writing" — collectively account for nearly 80% of all conversations. Writing is particularly dominant for work tasks, underscoring chatbots' distinctive advantage over traditional search engines in generating digital outputs. Notably, computer programming and self-expression represent relatively small shares of overall use, challenging common assumptions about LLM chatbot adoption patterns.
The paper concludes that ChatGPT's primary economic value lies in decision support, which is especially impactful in knowledge-intensive jobs. The study was conducted with Harvard IRB approval, though several authors are OpenAI employees, a disclosed affiliation.
