On the Nature of AI Development Processes: Between Recipes and Cooking, Bridging Static Models and Dynamic Enactments

Abstract

AI-enabled systems emerge from interdependent software, data, and machine-learning processes, yet their development is often organized through fragmented practices and tools rather than an explicit, coordinated engineering process. We argue that this fragmentation reflects two related gaps: the absence of a process structure spanning ML and non-ML development, and a disconnect between process specification and enactment. To address these gaps, we propose AIEngOrchestrator, a model-driven framework that connects structural process definition with executable enactment. First, we extend the Rational Unified Process (RUP) to define a coordinated AI development process capturing cross-domain roles, activities, artifacts, and their dependencies. This process model then serves as the structural input to GSM4SE4AI, an executable artifact-centric domain-specific process modeling language developed following Model-Driven Engineering principles, which governs artifact lifecycles and their coordinated evolution at runtime. We evaluate the feasibility of the proposed framework through a prototype implementation and representative AI development scenarios.