Revisiting Business Process Analysis Through the Lens of Large Language Models: Prompting Experiments With BPMN Process Serializations

Recent position papers have proposed that the traditional Business Process Management (BPM) lifecycle must be revisited considering generative AI advances, specifically by investigating how Large Language Models (LLMs) can assist various phases of the lifecycle. Inspired by that call to action, this paper reports on a series of experiments on how OpenAI's GPT-4 responds when querying the content of Business Process Model and Notation (BPMN) diagrams, as potential support for the Analysis phase of the BPM lifecycle. We are particularly interested in how BPMN content—typically available in enterprises that adopted the BPM lifecycle—should be exposed to LLM services, therefore we comparatively experiment with diagrams provided as XML serializations or as tool-specific RDF serializations. This is a comparison between a standard serialization characterized by intricate cross-referencing that compensates for the XML rigid hierarchical structure and the “semantic graph” view of RDF that is open-ended in terms of semantic annotation and can be serialized as sequences of statements that resemble natural language. The quality of the answers is assessed using the Retrieval Augmented Generation Assessment framework.