How AI Agents Reshape Knowledge Work: Autonomy, Efficiency, and Scope
A · Strong evidence A Harvard Business School economist and Perplexity researchers analyse matched sessions from Perplexity's own logs with a stated model and robustness checks; strong field data, but all from the vendor's product and its early power users.
Summary based on the full text·Original: about 80 minutes to read
In short
Jeremy Yang (Harvard) with Kate Zyskowski, Noah Yonack and Jerry Ma (Perplexity), arXiv 2606.07489, June 2026. The paper compares Perplexity Search, a conversational answer engine, with Perplexity Computer, an agent orchestrator, using the same users over 27 February to 27 May 2026. A simple task model predicts that agents, with higher delegation cost but lower cost per step, push people toward larger, higher-value tasks. The data agree: agents do more autonomous work, cut time and cost, and widen the scope of tasks attempted.
Key numbers
- 26 minutes vs 33 secondsautonomous machine work per session on Computer versus Search for near-identical first queries
- 269 to 36 minutestask completion time with Search alone versus Computer plus a human, on matched tasks
- 87% / 94%estimated reduction in time and cost with Computer versus a human using Search
- 1.3% vs 2.9%queries with medium-to-high next-turn dissatisfaction on Computer versus Search
- 2.40 vs 1.74O*NET knowledge domains required per query on Computer versus Search
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- Computer's cumulative queries reached 84 times the first-week total within three months. Largest use cases: Research and Analysis (25.8%) and Document and Asset Creation (18.6%).
- In 10,000 matched session pairs (cosine similarity above 0.99), Computer performed 26 minutes of autonomous work versus 33 seconds on Search, a 48x ratio.
- Follow-up queries on Computer shift from manual directives to verification and extension.
- Dissatisfaction on the next turn fell from 2.9% to 1.3% of queries, a 55% reduction.
- Completion time fell from 269 to 36 minutes; estimated time and cost fell 87% and 94%. A Search-aided professional would need to finish all manual steps in under 20 minutes to break even.
- Computer queries cross users' primary occupation more often (gap of 9 percentage points across all 8 occupation clusters).
- Complexity: 71% abstract non-routine tasks versus 53%; Create-level work 50% versus 26%; 51% need three or more knowledge domains versus 17%.
- 23% of Computer queries that map to an O*NET task statement involve a task never seen in the same users' Search queries.
How they got it
Matched-session natural experiment within users, LLM-based classification of queries against O*NET and Bloom's taxonomy, assumed human-equivalent times per tool step with sensitivity and breakeven analysis, cross-validated by an LLM-driven procedure and user interviews.
Limits
- Three-month early-adoption window dominated by power users and paying subscribers.
- Matching works only for Computer queries with close Search analogues.
- Time and cost estimates rest on assumed human-equivalent times; authors say magnitudes are approximate.
- Heavy reliance on LLM classification; only Perplexity usage is observed; three authors work for the vendor.
Why it matters for companies
The paper suggests the gain from agents is less about speed on old tasks and more about people taking on work outside their role. That has direct consequences for job design and team size: if one person can produce what used to need several specialists, coordination roles shrink and verification skills matter more. Budget the human review time; the authors count it as the agent's fixed cost.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org