Artificial intelligence agents are already being used primarily for cognitive work rather than novelty or automation, according to a large-scale study by Perplexity and researchers at Harvard University. The analysis found that 57% of agent activity is focused on cognitive tasks, with 36% devoted to productivity and workflow and a further 21% to learning and research.
The study, described as the first of its kind at scale, analysed hundreds of millions of anonymised interactions from users of Perplexity’s Comet browser and Comet Assistant. It set out to understand who adopts AI agents, how intensively they are used and which tasks people are willing to delegate in real-world settings.
Perplexity characterised the dominant pattern as a “thinking partner” use case, where agents support analysis, synthesis and decision-making rather than executing narrowly defined commands. Examples highlighted in the research include procurement professionals scanning customer case studies, students navigating complex course materials and finance workers filtering stock options and analysing investment information.
The data also shows a clear evolution in how people use agents over time. New users typically begin with low-stakes, exploratory queries but migrate towards productivity-focused tasks as familiarity grows. Productivity and workflow categories showed the highest retention, while early engagement in learning or research was strongly correlated with longer-term active use.
Related reading
- What you need to know about ChatGPT Health as a secure hub for personal health information launches
- PayPal plugs into Google’s agentic checkout as AI-driven commerce moves to scale
- LEGO Education launches AI and computer science classroom kits for younger students
Usage is concentrated among a relatively small set of professions. Six core occupations now account for around 70% of all agent activity. Digital technologists alone generate about 30% of queries, while marketing, sales, management and entrepreneurship roles show particularly high stickiness. Patterns differ by group: finance professionals dedicate 47% of their queries to productivity tasks, students allocate 43% to learning and research, and personal contexts make up more than half of overall query volume.
The researchers argue the findings provide empirical evidence that AI agents are moving beyond experimentation and becoming embedded in daily work. Perplexity said the durability of productivity use cases points to the emergence of a hybrid intelligence economy, where humans increasingly rely on agents to augment, rather than replace, complex cognitive labour as the tools mature and integrate more deeply into workflows.
The Recap
- Perplexity and Harvard published a large-scale agent usage study.
- Six occupations drive seventy percent of agent activity.
- Productivity and learning tasks account for most long-term use.