How AI Agents Are Changing the Way We Work

New research shows the next wave of AI expands both productivity and the scope of work.

For businesses, the most important question about agentic AI is whether it actually changes what work people can get done. The new working paper, “How AI Agents Reshape Knowledge Work: Autonomy, Efficiency, and Scope,” begins to answer that question. Co-written by HBS AI Institute affiliate Jeremy Yang and using data from Perplexity, the paper outlines the immense changes in store when execution moves from workers’ hands to a machine’s. 

Key Insight: Autonomy and Environment

“Over the last few years, frontier products have progressed from conversational assistants to copilots to agents” [1]

Yang and his co-authors delineate agentic AI along two vectors: autonomy and context integration. They define autonomy as the ability to plan and act with limited human intervention, and context integration as the ability to read from and write to a user’s digital environment (their computer, browser, and external connections and services). While LLM chatbots like ChatGPT have limited autonomy and context integration, agentic tools like OpenAI Codex, Claude Cowork/Code, and Perplexity Computer have a much higher level of capabilities in both areas. For example, with an “agent orchestrator” product like Perplexity Computer, a user names an outcome, and the system does the rest: searches, codes, builds, coordinates with external systems, and completes the job.

Key Insight: The Economics of a Task

“The model centers on a shift in the cost structure.” [2]

The co-authors also construct an individual-level task-based framework to investigate agentic systems more deeply by modeling work tasks as a number of discrete steps and tracking the cost to complete them, where cost refers to the full burden of resources (time, effort, and actual money). Picture each task as a stack of steps. Conversational AI (e.g. ChatGPT, Perplexity Search) has a low fixed cost because it’s easy to start: ask a question, get an answer. But it has a high marginal cost because every extra step still requires the user to decide what comes next: ask another question, interpret the response, and often do additional work manually. On the other hand, an AI agent carries a higher fixed cost because the user must specify the goal carefully and later verify the output, but a lower marginal cost per step, because the agent can plan and execute much, if not all, of the workflow autonomously. This tradeoff suggests that agents only pay off once a task is long enough to amortize the setup. Below a threshold, the chatbot wins; above it, the agent does. It also suggests that agents expand the frontier of tasks worth attempting, pulling users toward higher-value work they previously couldn’t afford. 

Key Insight: Inside a Natural Experiment

“We therefore exploit user-generated natural experiments in which the same user submits near-identical initial queries (first message in a session) to both products.” [3]

Evaluating the actual impact of agentic tools requires a rigorous scientific approach that moves past theoretical claims. Using three months of data from February to May 2026, the researchers found instances where the same person opened nearly identical requests in both Perplexity Search (Conversational AI) and Perplexity Computer (AI agent), yielding 10,000 paired sessions that hold the underlying task essentially fixed. To make sure each agent session genuinely executed work rather than just chatting, they kept only sessions that invoked at least one “do” tool: code, browser actions, file writes, or external API calls. They then estimated time and cost for a Perplexity Search + Human workflow in three ways: a tool-by-tool estimate and an independent LLM estimate that were both benchmarked against real wage data from the Bureau of Labor Statistics, plus interviews with 25 active users.

Key Insight: Agentic Supremacy

“If agents lower barriers to entry across occupational boundaries and expertise levels, and reduce coordination costs, individuals may produce outputs that previously required teams.” [4]

The findings of their empirical research reveal a massive shift in productivity. The Perplexity Computer + Human workflow (where a human supervised an agentic-executed workflow) reduced average task completion time from 269 minutes (of estimated Perplexity Search + Human completion time) to 36 minutes, an 87% reduction. This matched the 84% time savings found from an independent LLM-based estimate. Importantly, these gains did not appear to come at the expense of quality. For multi-turn interactions, the researchers looked at the user’s follow-up message after an AI response. If the next message included a correction, retry, re-ask, error report, or similar signal, it counted as dissatisfaction. Perplexity Computer produced lower next-turn dissatisfaction than Perplexity Search. 

Productivity gains are only part of the story: users’ roles and the scope of their work also changed. As execution was delegated to the agent, Perplexity Computer users shifted their time from doing the work to verifying and extending it. They also crossed occupational boundaries more often, taking on tasks outside their core expertise, attempting more cognitively complex work, and drawing on broader knowledge domains than the same users did with Search.

Why This Matters

This research suggests that AI agents can unlock value by helping humans push at the edges of what their roles can include. A manager may prototype a dashboard, a marketer could conduct deeper financial analysis, a lawyer might produce more polished client-ready materials, or a founder may move further into product and design work. For business leaders and executives, implementing AI adoption in your organization needs to include an expanded vision of your people’s potential as much as it includes increased technological productivity and efficiency. 

References

[1] Yang, Jeremy, Kate Zyskowski, Noah Yonack, and Jerry Ma, “How AI Agents Reshape Knowledge Work: Autonomy, Efficiency, and Scope,” arXiv preprint arXiv:2606.07489 (June 2026): 2. https://doi.org/10.48550/arXiv.2606.07489 

[2] Yang et al., “How AI Agents Reshape Knowledge Work,” 4.

[3] Yang et al., “How AI Agents Reshape Knowledge Work,” 15.

[4] Yang et al., “How AI Agents Reshape Knowledge Work,” 39.

Meet the Authors

Watch a video version of the Insight Article here.