How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations
Citation. Wang, Z. Z., Shao, Y., Shaikh, O., Fried, D., Neubig, G., & Yang, D. (2025). How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations. arXiv. https://arxiv.org/abs/2510.22780
Summary. Wang and colleagues directly compare human and AI-agent computer-use workflows across 16 long-horizon tasks spanning data analysis, engineering, computation, writing, and design. Human and agent workflows often contained similar ordered steps, but agents overwhelmingly executed those steps programmatically rather than through the UI-centric and perceptual methods humans used. Agents were much faster and cheaper yet produced lower-quality work, including failures involving fabricated data and tool misuse, making workflow strategy and verification burden central to human-agent task allocation.
Key claims
- Human and agent workflows shared 83.0% of steps, and the order of aligned steps was preserved 99.8% of the time, indicating procedural similarity despite different execution strategies.
- Agents used programming tools in 93.8% of workflow steps, including open-ended and visually dependent design work where humans relied more heavily on user interfaces and perceptual judgment.
- In the observed human workflows, selective AI augmentation improved efficiency by 24.3% with limited workflow disruption, whereas end-to-end AI automation slowed human work by 17.7% because verification and debugging added work.
- Agent outputs were lower quality than human outputs and included data fabrication, intent misinterpretation, vision failures, tool misuse, and difficulty translating between program- and UI-friendly formats.
- Depending on the agent configuration, agents completed work 88.3–96.6% faster and at 90.4–96.2% lower estimated cost than the average human worker, supporting selective delegation of readily programmable steps rather than an inference of end-to-end replacement.
Evidence & limitations
- The study compares 48 qualified Upwork professionals and four representative agent configurations on 16 curated computer-use tasks. The authors map the five sampled skill domains to 287 computer-using U.S. occupations and estimate coverage of 71.9% of their daily work activities, but the empirical task sample remains small relative to that claimed occupational reach.
- The workflow-induction toolkit converts heterogeneous mouse, keyboard, screen-state, and agent trajectories into hierarchical steps and uses automated alignment to compare workers; conclusions depend on the validity of those induced representations and task rubrics.
- The paper evaluates bounded computer-use tasks rather than implemented organizational workflows, so it does not establish durable productivity, job redesign, worker learning, accountability, or labor-substitution effects.
- This is arXiv version 2, a preprint updated November 6, 2025. Crossref returned no matching published record or non-arXiv DOI.
- The public arXiv v2 PDF exactly matches the ingested Drive PDF by SHA-256.