Stanford's WORKBank Reveals Where AI Should Automate or Support Jobs

AI Agents Transforming Job Execution

AI agents are revolutionizing how tasks are performed by executing complex, goal-oriented workflows. Unlike static algorithms, these agents use multi-step planning and software tools to manage entire processes across sectors such as education, law, finance, and logistics. Their adoption is already underway as workers leverage AI to assist with diverse professional responsibilities, reshaping human-machine collaboration.

Aligning AI Capabilities with Worker Preferences

A major challenge is the gap between what AI can do and what workers want AI to do. Even with capable AI systems, workers might resist full automation due to concerns about job satisfaction, task complexity, or the importance of human judgment. Conversely, some tasks workers are eager to automate lack sophisticated AI solutions. This mismatch is a barrier to responsible AI deployment.

Expanding Assessment Beyond Selected Roles

Previous AI adoption studies focused on limited roles like software engineering or customer service and emphasized company productivity over worker experience. These often analyzed current usage and lacked a future-oriented perspective, resulting in AI tool development that doesn't fully consider worker needs and preferences.

Introducing WORKBank: Survey-Based Worker Insights

Stanford researchers developed WORKBank, a dataset combining survey responses from 1,500 workers and evaluations from 52 AI experts using data from the U.S. Department of Labor’s O*NET. Through audio-supported mini-interviews, nuanced worker preferences were captured. CENTRAL to the framework is the Human Agency Scale (HAS), a five-level metric measuring the desired degree of human involvement in tasks.

Human Agency Scale (HAS): Balancing AI and Human Control

The HAS ranges from H1 (full AI control) to H5 (complete human control). Not all tasks benefit from full automation; for example, data transcription or routine report generation (H1 or H2) suit AI execution. In contrast, tasks like training program planning or security discussions (H4 or H5) require strong human oversight. Workers rated tasks on their automation desire and preferred HAS level, while experts assessed AI capabilities.

Insights from WORKBank: Acceptance and Resistance Patterns

WORKBank shows that about 46.1% of tasks have high worker desire for automation, mostly repetitive or low-value tasks. Tasks involving creativity or interpersonal interaction face strong resistance despite AI capability. Tasks cluster into four zones: Automation Green Light (high capability and desire), Automation Red Light (high capability, low desire), R&D Opportunity (low capability, high desire), and Low Priority (low desire and capability). Notably, 41% of tasks in Y Combinator-backed companies fall into Low Priority or Red Light zones, indicating potential misalignment with worker needs.

Implications for Responsible AI Workforce Integration

This research provides a framework that integrates technical feasibility with human values, offering actionable insights for AI development, labor policy, and workforce training. It highlights the importance of considering worker preferences to responsibly implement AI automation and augmentation.

For further details, check out the full paper and follow related discussions on social media and forums.