How AI Agents and Humans Approach Professional Work Differently—Evidence and Strategies for Designing Effective Human-Agent Systems
Citation. Westover, J. H. (2026). How AI Agents and Humans Approach Professional Work Differently—Evidence and Strategies for Designing Effective Human-Agent Systems. AI at Work: The Journal of AI, Organizational Change, & Workplace Transformation, 1(1). https://doi.org/10.70175/aiatwork.2026.1.1.3
Summary. Westover synthesizes a comparative study of human and AI-agent computer-use workflows across data analysis, engineering, computation, writing, and design. The article argues that agents tend toward programmatic execution even for visually dependent work, can produce lower-quality outputs masked by fabrication or tool misuse, and nevertheless complete the studied tasks substantially faster and at lower estimated cost. It frames task programmability, human verification, workflow-informed agent training, and hybrid task allocation as organizational design responses.
Key claims
- The article reports 83.0% human-agent workflow-step overlap and 99.8% order preservation in the underlying study, with lower alignment for open-ended design tasks.
- Agents reportedly used programming tools in 93.8% of workflow steps, contrasting with humans’ more UI-centric and perceptual approaches.
- Agents reportedly completed the studied tasks 88.3% faster and at estimated costs 90.4–96.2% lower, while success rates lagged human performance and failures included fabricated data, computational errors, and tool misuse.
- Selective AI assistance preserved more similarity to independent human workflows than end-to-end automation, which introduced verification, debugging, and error-correction work.
- The article recommends assigning programmable components to agents while preserving human quality assurance, contextual judgment, and workflow-level accountability.
Evidence & limitations
- This is a secondary synthesis, not the original empirical report. Its quantitative claims come from 2025-wang-ai-agents-human-work, the primary Wang et al. preprint How Do AI Agents Do Human Work? (arXiv:2510.22780); the two records are one evidence lineage, not independent corroboration.
- It provides no independent dataset, methods, or analysis through which those claims could be evaluated separately.
- Crossref identifies the canonical record as a March 2026 article. The PDF contains an older, nonresolving 2025 DOI string and early-access date.
- The official publisher PDF exactly matches the Drive copy by SHA-256.
- The journal record is verified, but the dry-run did not independently establish peer review;
publication_statustherefore remains conservativelyother.