AI in the Loop: How AI Affects Roles, Collaboration, and Trust in Agile Software Development

Abstract

Large language model-based tools have rapidly become part of the day-to-day workflow in software engineering, with growing evidence that they affect quality and productivity on an individual level. Less, however, is studied about how these tools shape collaborative practices within agile and DevOps teams, where communication, shared responsibility, and iterative decision-making form the basis of effective work. This thesis investigates the perceived effects of these tools on decision-making, role distribution, collaborative practices, and trust in AI-generated code within agile and DevOps team environments. A mixed-methods approach is used, combining semi-structured interviews with eleven developers across a varying range of company sizes, sectors, roles, and experience levels, and a complementary survey. The data collected through the interviews were analyzed using Braun and Clarke’s thematic analysis framework, supported by Socio-Technical Systems theory and Lee and See’s framework for trust in automation. The findings indicate role changes are ongoing rather than fully completed, with participants describing AI tools being used for decision-making rather than as delegated decision-makers. AI was described as displacing informal channels of knowledge transfer through which knowledge has historically been shared within agile teams. This displacement was particularly found between junior and senior developers. AI was also described as an amplifier of existing quality practices rather than a transformer. Trust was actively constructed through manual verification, such as review and testing, with experienced users reporting more calibrated than uniform trust. The study contributes to prior work by examining AI’s effects at the team level, a largely overlooked area in the literature focused on either the individual developer or the organization.