AI-Driven Automation of Backlog Item Generation: Automating Translation of Change Requests into Backlog Work Items

Sammanfattning

Developments in recent years have seen Artificial Intelligence (AI) and Machine Learning (ML) increasingly integrated within organizations and industries to harness their potential in automating tasks. Today’s modern agile framework emphasizes efficient working methods, reduced waste, and continuous improvement, which AI can strengthen through its ability to analyze large amounts of data and automate repetitive tasks. AI is today integrated into agile environments to refine the way of working, for example, sprint planning and backlog management. The main question this research aims to answer is “How can an AI based application be designed and implemented to effectively translate change requests into backlog items and fulfill the defined requirements, in a software development context?”. To answer the research question and develop an artefact, Design Science Research (DSR) will be applied, and all the steps will be followed. The steps mentioned are to explicate the Problem, Define Requirements, Design and develop the artefact, Demonstrate the artefact, and Evaluate the artefact. Using DSR, a method was developed to convert change requests into actionable backlog items using a Large Language Model (LLM). The method includes input classification, clarification questions, generation of acceptance criteria, code impact assessment, and the creation of the backlog work item. The method was developed and evaluated as a tool in realistic scenarios to test its relevance. The Design Science Research methodology, through its iterative framework, proved very productive and useful in producing a methodological artefact. Through continuous meetings and interviews with experts in the field, the problem description and requirements could be created, the artefact designed, and finally evaluated. This thesis proposed an artefact that automates the conversion of change requests into actionable backlog items. The artefact is described in a method but is tested as a tool within the organization. Future research can compare the performance of different LLM models, create a custom model with domain-specific data, and evaluate the artefact in a realistic environment.