One of the main tasks of new-generation cellular networks is the support of the wide range of virtual services that may be requested by vertical industries, while fulfilling their diverse performance requirements. Such task is made even more challenging by the time-varying service and traffic demands, and the need for a fully-automated network orchestration and management to reduce the service operational costs incurred by the network provider. In this paper, we address these issues by proposing a softwarized 5G network architecture that realizes the concept of ML-as-a-Service (MLaaS) in a flexible and efficient manner. The designed MLaaS platform can provide the different entities of a MANO architecture with already-trained ML models, ready to be used for decision making. In particular, we show how our MLaaS platform enables the development of two ML-driven algorithms for, respectively, network slice subnet sharing and run-time service scaling. The proposed approach and solutions are implemented and validated through an experimental testbed in the case of three different services in the automotive domain, while their performance is assessed through simulation in a largescale, real-world scenario. In-testbed validation shows that the use of the MLaaS platform within the designed architecture and the ML-driven decision-making processes entail a very limited time overhead, while simulation results highlight remarkable savings in operational costs, e.g., up to 40% reduction in CPU consumption and up to 30% reduction in the OPEX.