Details
- Perplexity announced new joint research with Harvard examining how work shifts from chat-style interfaces to autonomous agents, focusing on its Computer product.
- Over a three-month period, workers using Computer to execute tasks completed them in 87% less time and at 94% lower estimated cost than when using Search alone, while reporting higher satisfaction.
- The study constructed 10,000 matched pairs of sessions where the same user issued near-identical queries to both Search and Computer, with Computer required to run at least one execution tool; human time was monetized using U.S. Bureau of Labor Statistics wage data.
- The research models a cost structure where Search is cheap to start but requires humans to perform each step, while Computer has a higher upfront delegation and verification cost but much lower per-step cost as the agent executes the work autonomously.
- Results show Computer queries compounding over the study, reaching 84x their first-week volume, with most usage in knowledge work, especially research and analysis (25.8%) and document or asset creation (18.6%), concentrated in software, finance, and marketing domains.
- On a per-session basis, Computer runs about 26 minutes of machine execution compared with 33 seconds for Search, and 7.9% of Computer sessions call external connectors (averaging 15 calls per such session) versus 1.8% for Search.
- Computer queries were nearly three times as likely as Search queries to span three or more expertise areas, such as law, finance, or biology, and 23% of Computer work involved tasks that did not appear in the same users' Search histories.
- The paper finds that greater autonomy in agents like Computer correlates with higher output quality and user satisfaction, while also expanding the scope and complexity of tasks that users are willing to attempt.
- Perplexity frames the findings as evidence that autonomous agents can boost autonomy, improve quality, cut time and costs, and broaden the frontier of economically viable knowledge work tasks.
- The full research paper is available via the linked publication page.
Impact
This study gives empirical backing to the thesis that autonomous AI agents can materially change the economics of knowledge work by shifting effort from manual execution to oversight. By documenting large time and cost savings on real user tasks, Perplexity strengthens the competitive case for agentic systems against traditional chat assistants, and adds academic credibility through its collaboration with Harvard, likely influencing how enterprises and policymakers think about AI-driven productivity gains and task design.
