Managing risks associated with the evolution of systems with AI Components

Sammanfattning

As more companies add AI components to their software, most need to consider how these applications will evolve in the long term. The evolution of these AI components comes with risks, some of which come from the engineering practices adhered to when they are built. This is because traditional software practices do not apply when building AI applications. This creates technical debt, eventually impacting projects when project managers do not invest in mitigating them. This thesis discusses technical debts and describes the risks stemming from them. Mitigating these technical debts is a way to manage the risks to which these technical debts expose AI components and the larger system. Therefore, the research questions are: What risks from technical debt are associated with evolving AI components in software systems? What strategies are software companies adopting to assess risks associated with the evolution of AI components in software systems? And what approaches do software companies take to mitigate technical debt risks associated with the evolution of AI components in software systems? To answer these questions, we employed the qualitative analysis strategy. The data was collected by interviewing Software Engineers, Data Scientists, AI Engineers, Machine Learning Engineers, and Engineering Managers from various companies using predetermined questions. The analysis was done using the thematic analysis method. We found that the risks in the literature varied from those reported by our respondents. We also discovered two risk categories: system and business risk. We discovered that all respondents assess risk via monitoring and that risk mitigation varied slightly in literature from what our respondents reported. In conclusion, we corroborated the risks and mitigation strategies identified in the literature with our respondents, noting some differences. We learned that the primary assessment method used by all our respondents was monitoring and mitigation, which is crucial to prevent AI component failure.