From Requirement Text to Diagrams : Using Generative AI for UML Modeling in Industrial Software Engineering
Payel Mahapatra · Diva portal (Dalarna University Library) · 2026
Background. Modeling of requirements plays a crucial role in reinforcing requirements traceability within software design. By transforming textual requirements into structured visual representations, it enhances clarity, improves communication among stakeholders, and supports a deeper understanding of system behavior and dependencies. Analysing requirements through converting requirements into design diagrams remains a largely manual and a time-consuming process especially in large software development companies, even though there have been attempts made in using Generative-AI to automate the creation of software architecture diagrams. Objectives. The objective of this study is to analyse the current industrial process of translating unstructured requirements to UML diagrams, identify challenges and find out to what extent this process can be automated. Our aim is to investigate how a Gen-AI based tool should be designed and used in an industrial setting for automating the creation of UML diagrams from unstructured requirements. Methods. We conducted a Design Science Research at a small sized project team in Ericsson AB. We studied existing design artifacts and conducted semi-structured interviews with experts in the team who perform manual tasks of modelling diagrams from unstructured requirements, to understand the current process and identify pain points. We then created a prototype to address these challenges and refined it iteratively based on expert feedback. We created a conversational Gen-AI agent that automatically creates UML activity and sequence diagrams from available information sources. Results. The results show that project members rely on multiple sources of requirements, which they must interpret and combine to create UML diagrams. Gaps or incomplete information are typically resolved through workshops, and there is no standardized modeling tool in use, highlighting an opportunity for automation. Gen-AI tools can support this process, but the diagrams they produce must be trustworthy, easy to read, follow organizational standards, and avoid hallucinated information. Such tools should also prompt users for missing details and seek confirmation before finalizing the diagrams. Conclusions. Our research shows that fully autonomous diagram generation is neither practical nor desirable. Instead, tools should serve as intelligent assistants that enhance the efficiency of human experts. Ultimately, successful adoption of Gen-AI based modelling tools depends on designing tools that are trustworthy, transparent, and aligned with existing engineering practices.