Study reveals how people use AI agents in 2025 | Keryc
Perplexity and a Harvard team published the first large study on real-world use of AI agents. They analyzed hundreds of millions of anonymous interactions from users of Comet and Comet Assistant to answer three questions: who adopts these agents? how intensely do they use them? and what tasks do they delegate to their assistant?
The results that really matter
The main surprise isn't that people use agents for simple tasks, but that most use them for cognitive work. The study shows that 57% of agent activity focuses on work that requires thinking: gathering information, synthesizing, analyzing and helping make decisions.
36% of the most common tasks are productivity and workflow.
21% are learning and research.
Can you imagine asking an assistant to draft the first version of a financial analysis, review client cases to identify relevant uses, or guide you through a course syllabus? That's already happening. In the study examples, procurement professionals used the assistant to filter use cases before talking to a vendor; a student used it to navigate and analyze course materials; a financial analyst employed it to filter investment options.
Agents are functioning more like thinking partners than digital butlers.
From casual trial to indispensable tool
Usage behavior changes over time. On day one, many users make low-risk queries: travel recommendations, movies or curiosities. But with use, there's a clear shift toward productivity.
The transition is notable: someone who starts asking for vacation ideas can end up using the agent to debug code, summarize financial reports or prepare presentations. Productivity and workflow categories show the highest retention rates. Also, people who use the agent for learning or research early on tend to become long-term active users.
This mirrors the evolution of the personal computer: it began as a toy or household gadget and became indispensable because of spreadsheets and word processors. AI agents seem to be following a similar path.
Who adopts and who actually depends on the agent
Adopting an agent doesn't always mean using it intensively. The study distinguishes adoption from intensity, and that's where the real picture appears: six occupations account for 70% of all agent activity.
Digital technologists generate 30% of the query volume.
Fields with high knowledge intensity — marketing, sales, management and entrepreneurship — show greater "stickiness": when they integrate the agent, they use it much more.
Context matters: finance dedicates 47% of its queries to productivity; students allocate 43% to learning and research. In design and hospitality, use is strongly tied to concrete needs: media work for designers or trip planning for hospitality staff.
What this means for work and learning
The data provides evidence we're moving toward a hybrid intelligence economy. If most interactions are cognitive work, agents are scaling human intellectual capacity: they research, synthesize and prepare material so you make the final decision.
The question stops being whether people will use AI agents. The question now is how fast the rest of the economy will incorporate these assistants as steady companions for work, learning and creation.
Final reflection
If you use or plan to use an agent, don't see it only as a way to save minutes. Think of it as a partner that helps you think farther and faster. The key is to start with small tasks and let that relationship evolve: what begins as curiosity can become a tool that improves your work and your decisions.
