Application of Agentic AI and Large Language Models in Industrial Requirements Engineering : AI AND LLMs IN INDUSTRIAL REQUIREMENTS ENGINEERING

Sammanfattning

Industrial Requirements Engineering (RE) in safety-critical sectors moves through several man- ual document tiers before reaching code. Each tier transition is detailed, formal work that takes time and reviewer attention to maintain traceability. The research literature shows that very few pub- lished Artificial Intelligence (AI) and Large Language Model (LLM) systems for RE have reached production use. This thesis explores how AI and LLMs can support industrial Requirements En- gineering, in partnership with Westermo Network Technologies AB. Westermo is a Swedish man- ufacturer of industrial data communications equipment for mission-critical sectors such as rail, energy, and water infrastructure. Design Science Research (DSR) is used as the methodology. Qualitative discussions with senior Westermo staff pointed to the manual step from the Market Requirement Specification (MRS) to the Functional Specification (FS) as the place where AI support would help most. In response, a four-agent pipeline is proposed. It is built around a shared Retrieval-Augmented Generation (RAG) knowledge base and two mandatory Human-in-the-Loop (HITL) checkpoints. A working prototype runs the pipeline end-to-end on three deployment backends. Three evaluation strands are used: a side-by-side comparison of five LLMs, a fault-injection test, and an expert review with two Likert forms and free-text comments. A model running locally came within one percentage point of a large cloud model on semantic similarity, and still produced a complete and fully traced FS. Defects that show up in the text are caught in every run on the cloud model, while missing content is caught in only one run out of five. The expert reviewers rated the conceptual design at a median of 4/5 and gave the symmetric review setup the maximum score. The novel multi-agent Requirements Engineering design shows that confidentiality and the human-oversight rules of the EU AI Act can be met end-to-end without losing output quality. Placing a dedicated Review Agent after every generative phase is a viable extension of existing multi-agent RE frameworks.