Booking a work trip used to involve juggling multiple tabs, prices, and flight times. Now you can tell an artificial intelligence agent what you want—a flight from Boston to San Francisco next week, ideally in the morning—and let the system handle the rest. This represents the shift to agentic AI, moving from asking AI for answers to delegating the entire task.
“It takes objectives and preferences, instead of step-by-step instructions,” says economist and data scientist Jeremy Yang, an assistant professor at Harvard Business School. “You can treat it as a personal assistant to take care of the lower-level details with little supervision for a wide range of digital work. There’s a lot of interest in agents based on how fast their capabilities are evolving and their potential downstream economic impact.”
Indeed, as companies increase adoption, Precedence Research estimates the global agentic AI market size will grow from $8 billion in 2025 to $199 billion by 2034, and PwC forecasts that its economic contribution could reach as high as $4.4 trillion annually by 2030.
To understand who’s currently using AI agents and how, Yang studied hundreds of millions of anonymized user interactions. His working paper, “The Adoption and Usage of AI Agents: Early Evidence from Perplexity” finds a clear pattern: The heaviest users tend to be knowledge workers, and they’re mostly using agents to boost productivity or assist with learning. Yang coauthored the paper, released in December, with a team from Perplexity that includes cofounders Johnny Ho and Denis Yarats; Jerry Ma, vice president of global affairs; data scientist Noah Yonack; and Kate Zyskowski, head of UX research.
Who’s using agentic AI?
The researchers analyzed data from Comet, an AI-powered browser developed by Perplexity, and its integrated agent, Comet Assistant. AI agents have three key traits: they are goal-oriented; act on external environments, such as files, apps, and websites; and operate with autonomy to complete tasks.
The researchers found that some of the heaviest users were:
Early adopters of the technology.
Users in higher-GDP countries.
Users in countries with higher education attainment.
Knowledge workers. Those in digital technology represented the largest career cluster, making up 28% of adopters. Academics and financial workers represented 10%, while those in marketing, design, and entrepreneurship accounted for 5%.
The top uses of agentic AI
The researchers found that queries clustered into a few top categories, ranked in this order:
Productivity. The largest category was related to productivity and workflow, accounting for 36% of total questions. This may include document and form editing, account management, email management, and spreadsheet and data editing.
Learning tasks. Courses and research, encompassing tasks like watching class videos and summarizing key content, accounted for 21% of total queries.
Media and entertainment. Movies, music, podcasts, and viewing and posting to social media comprised 16% of queries.
Shopping and commerce. Searching for and buying goods and services represented 10%.
Travel and leisure. Booking flights and hotels and creating trip itineraries totaled 7%.
Career-related tasks. Searching for and filling out applications for jobs came to 7%.
Much of the work was cognitively loaded, such as researching and summarizing key findings—less like an executive assistant and more like a research assistant, Yang says. “It’s like having a second brain and pair of hands,” he says.
The future of agentic AI at work
The researchers say their findings offer an early understanding of how agentic AI and the ecosystem around it might evolve.
Yang suggests businesses could use these insights to strengthen certain agentic capabilities—or even redesign digital interfaces for a world where AI becomes a primary user.
“If you’re a website and every click is coming from an agent, then you might design your interface in a different way than you would for a human,” Yang says.
At the same time, uneven adoption could widen existing disparities if left unaddressed, and human oversight will remain critical, especially for high-stakes and irreversible tasks. Yang also sees a role for policymakers in shaping regulations and for educators in preparing students for an AI-enabled workplace.
As agents advance quickly, he notes, they are better understood as digital coworkers than as tools, and employees may soon find themselves working alongside them.
“It sounds like science fiction, but the trend is clear,” Yang says. “We need to learn how to develop new skills and processes in order to be prepared.”
Illustration by Ariana Cohen-Halberstam with asset generated by Midjourney, an AI image generation tool.
