Introduction
Process modeling serves as a critical foundation for business analysis and organizational optimization. However, the creation of formal models, such as those using Business Process Model and Notation (BPMN), typically requires specialized expertise and significant manual effort to translate ambiguous natural language descriptions into rigid logical structures. While Large Language Models (LLMs) like GPT-4 have shown promise in text transformation tasks, they often struggle with the "structural integrity" required for formal modeling, frequently producing models with logical inconsistencies or unreachable states.
ProMoAI is introduced as a novel framework to bridge this gap, enabling non-experts to design and refine complex business workflows through simple natural language interaction while maintaining strict formal logic.
Methodology: Formal Guardrails and Iterative Refinement
The ProMoAI system architecture is designed as a multi-stage pipeline that treats process modeling as a code generation and execution task. This approach allows the system to leverage standard software debugging techniques to ensure model validity.
Advanced Prompting and Data Leveraging
To guide the LLM, ProMoAI employs several advanced prompting strategies. Role prompting establishes the AI as an expert process analyst, while negative prompting explicitly instructs it to avoid common logical pitfalls. Furthermore, the system utilizes few-shot learning by providing high-quality examples from the PET (Process Extraction from Text) dataset, which helps the model understand the mapping between textual descriptions and process structures.
The POWL Intermediate Language
A core innovation of ProMoAI is the use of the Partially Ordered Workflow Language (POWL) as an intermediate representation. Instead of generating BPMN directly—which is prone to structural errors—the LLM generates Python code that utilizes a library to construct a POWL model. POWL is a hierarchical modeling language that inherently prevents "dead parts" and ensures reachability, providing a mathematical guarantee of model soundness.
Secure Execution and Self-Healing
The generated Python code is executed within a secure sandbox. If the execution fails or validation rules are violated, the resulting error messages are fed back into the LLM. This creates a "self-healing" loop where the AI attempts to correct its own code based on technical feedback before presenting a final model to the user.
Results: Modeling the Online Shopping Journey
The capabilities of ProMoAI were evaluated using a complex online shopping scenario involving login, item selection, payment method setup, and delivery/return logic. The system successfully modeled various control flow constructs, including sequences, concurrency, and loops.
A key finding was the effectiveness of the Human-in-the-Loop approach. In the demonstration, the user provided iterative feedback—such as allowing the user to skip reward selections or modeling item selection as a repeatable activity. ProMoAI successfully integrated these refinements into the conversation history and updated the formal model accordingly. The final outputs are fully compatible with industry standards, exportable as BPMN or PNML files for use in existing Business Process Management (BPM) suites.
Conclusion and Future Directions
ProMoAI demonstrates that combining LLMs with formal intermediate languages and iterative feedback loops can democratize process modeling. By using code generation as a bridge, the system enforces structural constraints that LLMs cannot maintain on their own.
Future research will focus on making the system more LLM-agnostic, as current performance is optimized for GPT-4. Additionally, the authors plan to explore direct BPMN generation with enhanced validation techniques to allow for richer annotations while maintaining the logical rigor provided by the current POWL-based approach. Addressing the inherent stochasticity of LLMs to ensure consistent visual layouts across different runs remains another area for improvement.
Summary updated