Business Process Management (BPM) plays a crucial role in organizations of all sizes. An essential part of BPM is business process modeling to make the internal knowledge of processes explicit and create a shared understanding of how the process is executed. However, this is a very time-consuming task requiring a fundamental understanding of the underlying modeling language and the application of internal style guidelines to establish correct, clear, complete, and consistent models. This has become a common struggle for many companies. With the emergence of large language models (LLMs) and their remarkable ability to solve language reasoning tasks, knowledge workers across various industries have the potential to enhance their productivity. Recent research has also explored the application of LLMs in generating business process models. However, there is limited knowledge about incorporating the aspect of standardization by generating models adhering to internal style guides.

To address these challenges and offer a practical solution, this master thesis adopted the research strategy of design science research. By conducting insightful interviews with process engineers at cablex AG, a profound understanding of their current practices and style principles was gained. An analysis of the encountered process modeling challenges validated the issues identified in existing literature. Furthermore, an interview with the Chief Information Security Officer at cablex shed light on the security requirements associated with utilizing LLMs for handling internal knowledge, including considerations such as General Data Protection Regulation compliance and data protection. Equipped with this comprehensive understanding, an LLM application prototype was developed utilizing OpenAI's GPT-4 API, employing expert and few-shot prompting techniques. This artifact empowers non-BPMN experts to effortlessly generate business process models adhering to internal style guides through a user-friendly interface, harnessing the capabilities of LLMs.

The resulting output is conveniently presented in an interchangeable XML BPMN format, allowing for easy downloading and immediate visualization in the widely recognized BPMN 2.0 notation within the application. To evaluate the quality of the generated models, an assessment was conducted using the PET dataset. The results demonstrated that the LLM application performs comparably to human process model experts in terms of recall and precision. Moreover, a qualitative evaluation utilizing the system usability score revealed its above-average usability, exceeding an established reference benchmark. By addressing real-world problems and bridging gaps in existing research, this thesis contributes to the advancement of style guide-compliant business process modeling utilizing LLMs. Notably, the developed prototype and evaluation methodology offers practical insights for both practitioners and researchers and opens the way for improved process modeling practices in organizations.