With decentralized Machine Learning (ML) strategies and modern edge Tensor Processing Unit (TPUs), smart devices are no longer only consumers but also producers of Artificial Intelligence (AI), transforming an Internet of Things (IoT) into a global and decentralized Artificial Intelligence of Things (AIoT). With the availability of a large amount of AI actors comes not only the challenge to discover and network with them, but also the potential to use their AI capabilities as a service. However, the heterogeneity of the AI actors, their AI capabilities, AI contextual environment, mobility or even the AI characteristic available or sought requires not only a robust IoT architecture but also flexible AI semantics. In this paper, we present an AI-as-a-Service platform assisting AI consumers to identify existing AI tailored to their needs among the AIoT. We describe the architecture, the APIs, the message flows and AI semantics to identify the most appropriate AI workers when and where needed to efficiently generate AI models from distributed vehicles. As a proof-of-concept, we select an application scenario demonstrating the trainability/changeability of AI models between vehicles according to their context using the CARLA driving simulator.