Who Is Already Using AI Agents and Why They Find Them Indispensable?

A new study has examined, on a massive scale, how people interact with artificial intelligence agents in web browsers. The case study is Comet, Perplexity's AI‑powered browser, and its built‑in assistant, Comet Assistant. By analyzing hundreds of millions of anonymous interactions, researchers answered three key questions: who uses AI agents, how intensively they use them, and for what purposes. The findings reveal that early adopters are tech‑savvy users living in high‑GDP countries, with solid education levels, and working in digital or knowledge‑intensive fields such as technology, academia, finance, marketing, and entrepreneurship. To map the queries, a three‑level taxonomy was built: topic, subtopic, and task. The two biggest topics—Productivity & Workflow and Learning & Research—account for 57 % of all queries, while the subtopics “Courses” and “Shopping” make up 22 %. The top ten out of ninety tasks cover 55 % of requests. Personal use represents 55 % of queries, professional 30 %, and educational 16 %. In the short term users show strong stickiness to the assistant, but over time they shift toward more cognitively demanding topics, indicating that AI agents are becoming true thinking partners. These insights matter for researchers, businesses, policymakers, and educators, who must consider the impact of a rapidly evolving technology.

The analysis of Comet relies on an advanced data‑mining methodology that extracts usage patterns from an unprecedented data volume. Researchers anonymized interactions to protect privacy, applying aggregation and filtering techniques to remove identifiable information. The proposed hierarchical taxonomy draws from semantic ontology models, dividing requests into three levels: topic (e.g., Productivity), subtopic (e.g., Email Management) and specific task (e.g., Drafting a reply). This approach quantifies query distribution and pinpoints high‑interest areas. Additionally, the study employed "stickiness" metrics based on user return frequency and average session duration, revealing that users keep the assistant active longer when tasks demand complex reasoning. Another observation concerns the "cognitive shift": over time, requests move from operational tasks (e.g., bookings) to more analytical ones (e.g., summarizing scientific articles). This suggests growing trust in AI's reasoning abilities. Economically, the correlation between adoption and per‑capita GDP indicates that entry barriers remain tied to socio‑economic factors, yet the rapid spread of more capable agents could narrow these gaps. Finally, educational implications are significant: educational institutions will need to integrate AI agents into curricula, preparing students to collaborate with AI systems to enhance learning and productivity.