LLM-Assisted Requirements Engineering in Agile MDD: Industry Insights and Validation

Abstract

Despite the growing interest in integrating Large-Language Models (LLMs) within software development, there is limited empirically-grounded guidance for teams to effectively apply this technology in industry. We explore the use of LLMs for generating requirements artifacts within a low-code consultancy organization that builds low-code development applications following a custom Agile model-driven development (Agile MDD) process. Through the analysis of multiple project cases within the company, we construct a method as an accurate representation of the employed Agile MDD approach. We then identify high-potential use cases for LLM adoption within the Agile MDD method. For these use cases, we engineer reusable LLM prompts that generate requirements artifacts. We validate the generated output for three of such use cases: generation of user stories, of acceptance criteria, and of data models. The team members of four projects were asked to express their opinion on the automatically generated artifacts. The results show high appreciation for the artifacts, which were found to be mainly relevant and similar to requirements that were included in the initial specification. The generation of the data model was rated less positively than the other use cases. Besides providing detailed insights on the inclusion of LLMs in the company’s Agile MDD process, we share our results to provide guidance for other software teams seeking to leverage LLMs in Agile MDD.