Abstract:
Large language models (LLMs) are being rapidly adopted for complex knowledge work, yet little is known about how end users are adapting these tools to their needs, why their customised workflows take the shape they do, and what those adaptations reveal about unmet needs in human-AI interaction. This paper reports findings from seventeen semi-structured interviews with knowledge workers, primarily software engineers, investigating how complementary human-AI relationships develop in practice, and whether users perceive LLMs to be genuinely augmentative. Expert software engineers had developed workflows that preserved their agency over the parts of their work they valued. These were characterised by meticulous upfront specification and iterative agent configuration, which kept strategic and creative decisions in human hands. Even within these expert workflows, however, participants described new cognitive burdens, particularly regarding context management, output evaluation, and maintaining strategic oversight, all of which current tool design does little to mitigate by default. These findings suggest that complementarity is not a default property of LLM-based tools but something users must deliberately construct, often in ways those tools do not naturally support.
PPIG 2026 - 37th Annual Workshop
Agency, Expertise, and the Limits of Augmentation in LLM-Assisted Programming Workflows