You Shall Not Pass! Where and Why Developers Draw The Line on AI Autonomy

As generative AI tools become standard in software engineering, the question of how much autonomy to grant these systems has become critical. This study investigates where and why professional developers draw the line on AI autonomy. Drawing on cognitive appraisal theory and work design research, the authors surveyed 448 developers at Microsoft to understand their preferences across 20 software engineering tasks. The researchers classified these preferences using a five-level autonomy scale—ranging from no AI involvement to full automation—and modeled how task-specific appraisals (value, identity, accountability, and demand) and individual traits (AI experience, risk tolerance) predict the boundaries where developers cede control.

Most developers accept AI as a collaborator (e.g., generating code or tests) provided they retain final approval authority. However, they draw firm boundaries based on the nature of the work.

• Accountability as a Gatekeeper: When developers feel high accountability for a task, they are significantly less likely to allow AI to act on their behalf. This represents the first major boundary (the action boundary), where the risk of automation error becomes personally attributable.
• Identity and Judgment: Tasks central to a developer's professional identity are rarely delegated. Even when developers accept AI-generated artifacts, they often refuse to relinquish decision-making authority for these tasks.
• The Role of Demand: Interestingly, while high task demand (workload) typically leads to burnout, it also increases the likelihood that developers will delegate decision-making to AI to cope with pressure.
• Individual Differences: Developers with higher AI experience and greater risk tolerance are more willing to grant AI higher levels of autonomy, suggesting that comfort with the technology and uncertainty plays a major role in these decisions.

The authors frame these findings as a series of cascading locks. If organizations allow tool defaults to dictate these boundaries, they risk creating "hollow" roles where developers lose the ability to exercise the judgment necessary for their own professional growth. By understanding these cognitive appraisals, managers and tool designers can create workflows that offload repetitive toil while protecting the high-value, judgment-intensive work that keeps developers meaningfully engaged.