AI brokers, generative AI programs that act autonomously to carry out duties, have repeatedly been promoted by tech firms as a ‘digital concierge’ able to making purchases, reserving journey, or scheduling conferences on behalf of the person, with out requiring their fixed enter.
However new knowledge means that this imaginative and prescient is in distinction to how persons are truly utilizing AI brokers.
Whereas these brokers are good at reserving lodges or dealing with rote chores, over 57 per cent of all agent exercise in Perplexity’s Comet browser was targeted on cognitive work, based on a brand new examine launched by Harvard College researchers in partnership with the AI search startup on Tuesday, December 9.
The examine, which researchers claimed was the primary large-scale effort to grasp how persons are utilizing AI brokers, discovered that 36 per cent of the commonest actions carried out by Perplexity’s AI browser agent had been productiveness and workflow duties, whereas 21 per cent of the duties needed to do with studying and analysis.
These findings are primarily based on the researchers’ evaluation of lots of of tens of millions of anonymised interactions from Comet and Comet Assistant customers. It was targeted on three elementary questions: Who’s adopting AI brokers? How intensively are they utilizing them? And what duties are they delegating to their AI assistants?
Notably, productiveness and workflow-related duties additionally had the best retention charges. Through the use of AI brokers for studying or analysis duties early on, persons are additionally more likely to turn into long-term energetic customers, as per the examine.
These findings come at a time when firms throughout industries need to undertake AI brokers. They problem some frequent narratives that would probably reshape how brokers may very well be used to be taught, work, and resolve issues.
Story continues under this advert
“This examine supplies the primary empirical proof that we’re shifting towards a hybrid intelligence financial system. The dominance of cognitive duties in our knowledge means that AI brokers are scaling human cognitive work. As these instruments mature, we count on the “stickiness” of productiveness duties to deepen,” Perplexity mentioned in a weblog submit.
Whereas 2025 was imagined to be the yr AI brokers took off, the surge in utilization or adoption charges has not but arrived. A number of specialists equivalent to Andrej Karpathy, a famed AI researcher, have additionally questioned the progress of agentic AI.
“I’m not truly certain who mentioned this however they had been alluding to this being the yr of brokers with respect to LLMs and the way they had been going to evolve. I used to be triggered by that as a result of there’s some over-prediction happening within the trade. In my thoughts, that is extra precisely described as the last decade of brokers,” Karpathy mentioned on a current podcast hosted by Dwarkesh Patel.
Who’s utilizing AI brokers? And the way?
Six core occupations drove 70 per cent of all agent exercise within the Comet browser, based on the examine. Whereas digital technologists led on quantity ( accounting for 30 per cent of queries), different fields like Advertising, Gross sales, Administration, and Entrepreneurship, demonstrated the best “stickiness”, as per the examine.
Story continues under this advert
Researchers additionally discovered that customers are deploying Comet Assistant, Perplexity’s AI browser agent, to unravel the precise friction factors of their trade. As an example, over 47 per cent of queries from finance professionals had been associated to productiveness duties. Queries from college students, then again, had been made up of 43 per cent of studying and analysis duties.
The examine additional highlighted a key development in person behaviour, the place individuals could begin off utilizing the AI agent for low-stakes duties like film suggestions or normal trivia and steadily shift to extra intensive productiveness duties. “A person would possibly begin by asking a few trip spot, however as soon as they use the agent to debug a Python script or summarize a monetary report, they not often return,” Perplexity mentioned.