Process maps are helpful tools for visualizing and understanding workflows (Zeitner & Peyinghaus, 2013), yet their creation remains labour-intensive and prone to inconsistencies, especially when dealing with unstructured data and complex processes. Traditional approaches often lack scalability and efficiency, hindering organizations from leveraging process maps for strategic decision-making and operational improvement (Brandl et al., 2008a).

A general challenge was the lack of literature, as this is a new field of research. An absence of standardized notations for process maps and a potential need for tailored AI implementations to address organizational specifics, were identified. Moreover, the limitations of AI in handling domain-specific knowledge were highlighted, what is reaffirming the importance of human oversight in ensuring the accuracy and relevance of generated process maps (Kampik et al., 2023). This thesis explored the use of artificial intelligence, specifically large language models, to automate the generation of process maps from unstructured enterprise documents. A dual-model evaluation was conducted using a cloud-based architecture, namely Copilot, and a local model, namely LLaMA 3.2 1B instruct, focusing on three key tasks: process identification, classification, and output structuring. To evaluate model performance, a manually curated ground truth was created for reference. The evaluation involved confusion matrix analyses, semantic matching, and prompt engineering iterations.

The findings revealed that while both AI systems offered value, the cloud-based model significantly outperformed the local model in terms of accuracy, contextual reasoning, and output usability. Copilot was able to extract implicit process knowledge and convert it into structured XML-based outputs, demonstrating its suitability for enterprise-level applications. However, the cloud setup introduced challenges related to data privacy and control. In contrast, the local model ensured full data sovereignty but lacked semantic depth and failed to deliver usable results in complex analysis tasks. Additional insights highlighted the importance of well-formulated prompts and document structure, as these factors had a direct impact on model reliability. Furthermore, process classification into management, core, and support categories and even their branch specific subcategories proved feasible for large language models. The study offers a practical evaluation framework and strategic recommendations for the implementation of AI-assisted process mapping, along with a portfolio matrix that links AI performance to strategic relevance. It concludes that hybrid AI architectures, combining local pre-processing with cloud-based semantic analysis, may offer a balanced solution for organizations seeking to leverage AI while maintaining control over sensitive data. By addressing a previously underexplored area in AI-supported business process management, this thesis contributes novel insights into how LLMs can be used for process discovery and for generating structured, organization-specific process landscapes from unstructured sources.