Microsoft study of 448 devs reveals strict boundaries for AI autonomy

A new mixed-methods study by Microsoft researchers (Choudhuri et al.) surveyed 448 professional developers to investigate where they draw the line on AI autonomy in software engineering. Drawing on cognitive appraisal theory, work design, and automation research, the study found that most developers accept AI producing work under their oversight, but tolerances vary widely across tasks and individuals. Acceptance was lowest for identity-defining work (tasks that make developers feel like 'real' developers), human-facing tasks (e.g., code reviews, client communication), and design-oriented work (architecture decisions). Developers with more AI experience and higher risk tolerance were more likely to grant AI decision-making autonomy.

Task accountability—being responsible for the outcome—was associated with lower odds of allowing AI to act on a developer's behalf. Task identity—the degree to which a task feels core to a developer's role—reduced willingness to grant AI decision-making autonomy. Conversely, high task demands (time pressure, complexity) increased developers' willingness to delegate decisions to AI. The findings suggest that preferences for AI autonomy reflect how developers cognitively experience their work. For tool designers and engineering leaders, this means preserving meaningful human work requires keeping humans in the loop for identity-critical, high-accountability tasks while automating repetitive, high-demand tasks where delegation feels safe.