Not yet human-reviewed
This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.
When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration
Citation. Wu, J., Ziems, C., Chan, R., Chen, N. F., Chua, C., Chung, G., Hahn, J., Lee, W. S., Liu, Z., Ng, J., Ong, D. C., Ong, J., Soon, D. R., Song, T., Tan, Z.-X., Tao, S., Yang, E., Yang, Y., Yin, S. X., Kan, M.-Y., & Yang, D. (2026). When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration. arXiv:2609.12482.
Summary. This CIVIC-AI workshop whitepaper proposes evaluating AI augmentation at the workflow level rather than through adoption, productivity, or headcount measures alone. It defines six conditions for genuine augmentation: durable net value, meaningful human control, clear accountability and recovery, deepening learning, preserved career pathways, and job purpose. A social-survey case study illustrates allocating bounded and reviewable execution to an agent while retaining human authority over purpose, validity, interpretation, and consequential use.
Key claims
- Apparent AI productivity gains do not establish durable net value unless evaluation includes human review, exception handling, rework, recovery, and cognitive burden shifted elsewhere in the workflow.
- Meaningful human control requires workers to have the competence, time, information, and authority to detect failures, challenge assumptions, constrain actions, override decisions, and continue work during system unavailability.
- Workflow evaluation should distinguish snapshot integrity conditions from longitudinal effects on learning, career pathways, agency, and job purpose because delegation can remove the formative tasks needed for later human oversight.
- Delegation is more appropriate when outputs are inspectable, errors reversible, and stakes bounded; humans should retain stronger authority over goals, validity, ambiguous evidence, and consequential use.
Evidence & limitations
- This is a 2026 workshop-derived whitepaper that advances a normative framework and one worked social-survey example rather than reporting a controlled empirical evaluation of the six conditions.
- The public arXiv PDF is text-equivalent to the ingested Drive PDF (same eight pages and normalized extracted-text hash) but has a different binary SHA-256; the record preserves the Drive artifact hash.
- The arXiv API returned HTTP 429 during retrieval, and Crossref exact-title search found no exact matching published record; public provenance is the stable arXiv identifier and public PDF.