Does personal preference shape how engineers use AI tools?

The authors' central claim is that agency over agentic AI in software work is set at the organization before individual preference enters. Their abstract says agency is "primarily constrained by organizational policies rather than individual preferences." The discussion supports this with what participants described: mandates, allow-lists, and security rules, such as a ban on non-company AI on corporate devices and a rule that confidential data not go to external services. Some employees were required to use a specific tool such as Cursor. The authors name the layer directly: agency is "preconfigured at the organizational layer (policies, tooling defaults, repos, CI guardrails) before individual preferences matter." The support is participant testimony from companies with different AI rules, not a measure of how much policy constrains compared with preference.

Within those bounds, the paper describes two routes. Experienced developers keep control "through detailed delegation," and the abstract credits seniors with using "pre-AI foundational instincts to steer modern tools." Novices, by contrast, "struggle between over-reliance and cautious avoidance." The debugging task gives one measured signal of what this costs. Six of 10 juniors reported lower confidence in coding without AI after the task (pre-task M = 3.90 ± 0.88, post-task M = 3.30 ± 0.82, Cohen's d = 0.71). The authors measured these with their own five-point self-ratings. From these observations they propose three practices: preserving individual agency through incremental changes, interrupting, and verifying output; evolving the mentorship pipeline so seniors pass on intuition and judgment; and prompt and code reviews, in which juniors document and justify key prompts while seniors oversee accountability.

Against the nearest notes, the paper's contribution is the level it locates the decision at. Does generative AI prevent juniors from getting entry-level work? describes work that never reaches the junior, and the policy finding offers a mechanism for that allocation, since the rules on permitted tools and tasks are fixed before any individual chooses. Does machine agency exist on a spectrum rather than binary? places a system on a scale of what the machine does; this paper's finding is that where a team sits on that scale is largely decided by its organization. The seniors' "detailed delegation" is a different axis from the one in Does vibe coding actually keep humans in the loop?, which contrasts task size. The excerpt does not compare the two.

What the excerpt does not establish is wide. Phase 1 interviewed five senior engineers who were all male, aged 28 to 46, and based in the US. Recruitment used convenience and snowball sampling, and the authors describe the work as a snapshot of the summer of 2025. The "primarily constrained" claim rests on what participants reported. The excerpt's limitations section stops after its first sentence, so no further caveats are visible. The confidence figures are self-ratings, so they show perceived change, not measured loss of skill. The implication is that company policy should be recorded as a variable in any study of how juniors and seniors use AI, and that the three practices are proposals to test, since the excerpt does not evaluate them.