Design implications of an AI-assisted error categorization tool in an industrial setting

Abstract

Implementing AI-powered tools into existing workflows to increase efficiency and remain competitive is a priority among many companies and organizations, especially in the tech industry. However, there is a potential gap in the research regarding how to implement such AI systems in a way that feels natural to the users already familiar with the existing workflows. Research shows that users confronted with an AI tool that upsets their workflow are less likely to trust the tool and provide feedback to it. This loss of feedback and trust in the AI tool can lead to decreased performance over time, potentially model collapse. In this thesis we explore potential solutions to this problem from an Interaction Design and UI/UX perspective, addressing a real-world problem in an industrial setting. The study was conducted in an industrial software development context at Ericsson, focusing on the CATegorizer, an AI tool that predicts problem types for failed software tests in the CI pipeline. A mixed-methods approach was utilized, with a substantial pre-study gathering both quantitative and qualitative data, followed by ideation, implementation and evaluation of the proposed solutions. The two implemented and evaluated solutions were a dashboard intended to increase understanding of the CATegorizer and its feedback loop, as well as an interaction mechanism for providing feedback more easily within the existing workflow. The results suggest that existing guidelines for Human-AI interaction provide a solid foundation for applications in an industrial setting, but need to adequately take into account the affected user group and the surrounding organizational and workflow context. For successful implementation of AI tools in professional workflows, it is important that the users are adequately informed of the feedback loop and the value of their feedback, while minimizing additional efforts required to provide it.