AI Agents Reshape Knowledge Work, Increasing Autonomy and Efficiency

According to the arXiv paper arXiv:2606.07489 by Jeremy Yang and coauthors, a Harvard Business School and Perplexity collaboration, the authors analyze production logs from Perplexity's Search and Computer products to compare conversational-assistant usage with agentic, autonomous workflows. The paper finds that Computer performs about 26 minutes of autonomous work per user session, versus 33 seconds for Search, measured on near-identical initial-query pairs. It reports per-query dissatisfaction rates 55% lower on Computer than on Search, matched-task completion time falling from 269 to 36 minutes, and estimated time and cost reductions of roughly 87% and 94% relative to working with Search alone.

The study uses matched-session natural experiments, pairing near-identical initial queries routed to each product to hold task intent roughly constant. The authors quantify shifts in follow-up query distribution, execution duration, and dissatisfaction. Per the paper, Computer queries more often cross occupational boundaries, bundle composite subtasks, and demand higher-order cognition, while Search users among the same population rarely attempt such tasks. The analysis is based on production telemetry rather than controlled lab tasks.

The results read as an empirical case study of large efficiency gains when autonomy handles multi-step, interdependent subtasks. A recurring pattern is that as an agent reliably performs decomposition and execution, user activity migrates toward verification, extension, and cross-domain synthesis, raising familiar questions about agent reliability, verification pipelines, and cost accounting for end-to-end automation versus human-in-the-loop work.

The reported magnitude of the time and cost reductions makes the study notable for practitioners tracking agent productivity. If similar effects replicate across platforms and domains, they would shift tooling priorities, benchmark design, and the engineering effort spent on integration and verification.

• •Independent replication on other platforms and domain verticals.
• •Transparent cost accounting that includes compute and human supervision.
• •Evaluation frameworks that measure verification workload, failure modes, and hallucination in composite tasks, not just single-turn accuracy.