Although the recent emergence of Large Language Models (LLM) has introduced new approaches to Entity Matching (EM), there has been limited effort towards combining diverse AI techniques and developing corresponding implementation frameworks. In a design-science-oriented way, this paper introduces a new kind of interplay between generative and restrictive AI for EM. This interplay is operationalized through the development of a novel orchestration framework and empirically demonstrated through the integration of three distinct AI types: (1) a fuzzy, restrictive Deep Learning-(DL)-based AI, (2) a fuzzy, restrictive Machine Learning-(ML)-based AI, and (3) a generative LLM-based Retrieval Augmented Generation (LLM-RAG). This real-world implementation demonstrated the efficacy of the framework. Findings show that the combination of (a) the modular software architecture having module layers, system layers, and usage layers and (b) the master-slave infrastructure model are suitable for the framework construction and enable enterprises to deploy containerized AI-modules independently or in orchestrated workflows. Each AI-module operates in isolated Docker containers, ensuring portability and scalability in heterogeneous and distributes hardware infrastructures. Furthermore, the AI interplay supports the combination of advantages of different AI techniques. Contributions include a reusable architecture for AI orchestration in further domains.