The Agentic AI Revolution Is Being Built By Individual Workers, Not Corporate IT
The enterprise AI story is being sold. The real one is being written by individual knowledge workers, one query at a time.
That is the counterintuitive picture emerging from the first large-scale field study of how people actually use AI agents in the wild. Harvard Business School economist Jeremy Yang and a Perplexity team analyzed hundreds of millions of anonymized user interactions with Comet, Perplexitys AI-powered browser, between July and October 2025. Their working paper, The Adoption and Usage of AI Agents: Early Evidence from Perplexity, published in December and resurfaced by HBS Working Knowledge this week, is the most granular look at real-world agent usage to date and it does not match the narrative being sold in conference keynote halls and earnings calls.
The story the AI industry tells is about enterprise transformation: CIO sign-offs, IT procurement cycles, company-wide rollouts. What the data shows is something more grassroots and more interesting. The heaviest agent users are not corporate IT departments running approved deployments. They are individual knowledge workers, often already in digital or technology roles, who discovered the tools on their own and never stopped using them. Early adopters made nine times as many agentic queries as users who came aboard after general availability. The first cohort did not just try the technology; they moved their work through it.
The dominant use case also subverts expectations. The conventional framing of AI agents as digital concierges, book my flight, order my lunch, handle the administrative friction, turns out to be roughly half the picture at best. Productivity and workflow tasks, 36 percent of queries, and learning and research, 21 percent, together account for 57 percent of all agentic activity. These are not offloading the boring stuff. They are using agents as research assistants and thinking partners: summarizing course materials, scanning case studies before a vendor meeting, filtering investment options and drafting analysis. A procurement professional used Comet to identify relevant use cases from customer documentation before negotiating with a vendor. A finance worker delegated the task of filtering stock options and pulling the underlying data. The agent handled information gathering and initial synthesis; the human made the final call. The workers are not avoiding work. They are amplifying it.
Who are these people? Digital technology is the single largest career cluster, at 28 percent of adopters. Academia and financial workers represent 10 percent; marketing, design, and entrepreneurship add another 5 percent. The typical user is someone who was probably already spending significant time in a browser, working with information. Which raises a question the paper acknowledges but does not resolve: is this finding about what agents are good for, or about who was already using Perplexity? The sample is self-selected. Enterprise users were excluded from the dataset. Workers in industries with less digital-native workflow may be adopting agents through completely different channels, or not at all.
The geographic pattern compounds the selection concern. Countries with higher GDP per capita and higher educational attainment show substantially higher adoption and usage intensity. The agentic AI revolution is arriving first for people who were already ahead: the same pattern that appeared with PCs, smartphones, and broadband internet, and that policymakers have spent two decades trying to close. If this data holds as agents move into broader deployment, the adoption gap is not a transitional artifact. It is structural.
For businesses, the implications cut both ways. On one side, individual workers adopting agents without IT approval creates exactly the shadow IT problem that security teams have fretted about since the first SaaS app slipped past the firewall. On the other side, if the most engaged agent users are already knowledge workers in high-GDP sectors, the enterprise sales motion, built around ROI calculations, security reviews, and multi-year contracts, may be selling to the wrong buyer. The agents are already in the building. IT just has not approved them yet.
The data also shows stickiness. Users who engaged with productivity, learning, or career-related agent tasks early were most likely to stay active over the study period. The 10 most common tasks out of 90 accounted for 55 percent of all queries, concentration rather than diffusion. The agentic AI market may be real, but it is being built around a relatively narrow set of high-frequency workflows before it broadens. That concentration is useful signal for anyone building or buying in this space. The current market is smaller than the forecasts suggest, and stickier than the adoption rates imply.
There are obvious caveats. The study is co-authored by Perplexity staff and uses only Comet browser data, a product and user base that skews young, male, and technically oriented. Three and a half months of behavioral data is early evidence, not proof of long-term adoption patterns. The query classification methodology, a human-labeled taxonomy of 90 tasks across three levels, is reproducible but not independently verified.
Those caveats do not kill the story. They sharpen it. The most interesting thing about this study is not any single number. It is the gap between the enterprise AI narrative being sold to investors, conference audiences, and IT buyers, and the ground-truth pattern of individual knowledge workers who found the tools, used them heavily for cognitive work, and did not stop. The agentic AI revolution may be real. It just is not arriving the way the vendors say it is.