Generative AI in Knowledge Work: Design Implications for Data Navigation and Decision-Making
Citation. Yun, B., Feng, D., Chen, A. S., Nikzad, A., & Salehi, N. (2025). Generative AI in Knowledge Work: Design Implications for Data Navigation and Decision-Making. CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3706598.3713337
Summary. Yun et al. study how knowledge workers synthesize unstructured information across tools and then prototype Yodeai, an AI-enabled system for product managers navigating user interviews, reviews, and related decision data. Across 20 formative interviews and a user study with 16 product managers, the paper argues that useful GenAI tools for knowledge work need adaptable user control, transparent collaboration and verification mechanisms, and context-aware interoperability with organizational data and workflows. The study is strongest as human-centered design evidence about knowledge synthesis and decision support, not as causal evidence that AI improves organizational decisions.
Key claims
- Knowledge workers reported recurring difficulty synthesizing unstructured information scattered across platforms, especially when they needed to organize materials, preserve context for collaborators, search beyond keywords, and prioritize evidence for decisions.
- In the formative study, 20 U.S.-based participants across journalism, startups, graduate work, consulting, product management, and related roles described needs for centralized structure, semantic search, thematic clustering, source visibility, stakeholder-aware communication, and integrations with existing work systems.
- Yodeai operationalized three AI-supported workflow templates for product-management knowledge work: Q&A with citations, Pain Point Tracker for clustering and quantifying issues over time, and User Insights for multi-granular summaries on a visual whiteboard.
- In a user study with 16 product managers, participants valued AI support for exploring large volumes of user feedback, but they also kept returning to source data, company context, collaborator judgment, and manual validation because product decisions were high-stakes and AI outputs alone were not trusted as decision authority.
- The paper identifies three design implications for GenAI in knowledge work: workflow flexibility and user autonomy, accountability and validation through multiple verification routes and audit trails, and secure interoperability with business context, priorities, and multiple data pipelines.
- The paper surfaces risks that map directly onto AI-mediated knowledge work: overreliance on AI, confirmation bias from agreeable systems, user isolation when AI use becomes solitary, data security and poisoning concerns, and contextual factors outside the model’s reach.
Evidence & limitations
- Evidence comes from qualitative semi-structured interviews with 20 knowledge workers, an All Our Ideas poll with 111 product managers, and a Yodeai user study with 16 product managers focused on user-interview/review analysis, feature and bug brainstorming, and prioritization.
- The sample and prototype are concentrated in software product-management contexts, so the claims generalize best to knowledge work that involves unstructured customer/user evidence, cross-functional collaboration, and high-stakes prioritization rather than to all professional work.
- The study captures short-term interaction with a prototype and self-reported work practices; it does not measure long-run adoption, productivity, decision quality, learning, or organizational outcomes.
- The source is a CHI 2025 paper with public DOI and arXiv provenance; the Drive PDF hash was computed during ingest.