Current and Future Use of LLMs for Knowledge Work

Analysis

Why This Paper Matters

This paper addresses a critical gap in the rapidly evolving landscape of Large Language Models (LLMs): understanding how knowledge workers actually use these tools in practice and what they envision for the future. While much research focuses on technical capabilities or benchmark performance, this work provides a human-centered perspective grounded in survey data from real users. The two-wave longitudinal design (surveys separated by a year) adds robustness and captures the dynamic nature of LLM adoption.

For AI practitioners and product teams, these findings offer actionable insights into the tasks that workers find most valuable today (e.g., code generation, text refinement) and the deeper integration they desire (e.g., LLMs operating on personal data and workflows). This can inform feature prioritization, user experience design, and deployment strategies for enterprise AI tools.

Technical Contributions

• Survey methodology with longitudinal validation: Two surveys (n=216 and n=107) conducted one year apart, allowing the authors to validate initial findings and track changes in usage patterns over time.
• Categorization of current vs. desired use: The study distinguishes between tasks workers already perform with LLMs and the more integrated, data-aware interactions they hope for in the future.
• Implications for design and adoption: The paper translates user-reported needs into concrete recommendations for building generative AI systems that better support knowledge work.

Results

The first survey (n=216) identified that knowledge workers primarily use LLMs for generating code and improving text. When asked about future desires, workers described a vision where LLMs are seamlessly integrated into their existing workflows and have access to their personal data. The second survey (n=107) confirmed these patterns and provided updated insights, suggesting that while core use cases remain stable, expectations for deeper integration are growing. No specific quantitative metrics (e.g., accuracy, productivity gains) are reported, as the study is qualitative in nature.

Significance

This research contributes to the growing body of work on human-AI interaction by providing a grounded, user-centric view of LLM adoption in knowledge work settings. It highlights the gap between current tool-level usage and the desired system-level integration, pointing toward future directions for generative AI design. For the AI field, it underscores the importance of studying real-world usage patterns to guide both technical development and organizational deployment of LLMs.