Recent years have witnessed a surge in automated machine learning (AutoML) tools and techniques. Such automation has been achieved through various techniques such as reinforcement learning-based agents, multi-arm bandits, numerical optimization methods such as Bayesian optimization and others. While the automation efforts proposed so far have led to efficiencies centred on the model building portion of data science, other areas are relatively untouched. For example, data engineering, which requires a semantic understanding of the data, remains manual to a large extent. AutoML techniques are also limited in the types of models tackled, for example targeting supervised models for forecasting, prediction, and classification, while stopping short of decision-making models. In this presentation, we will discuss the important shortcomings of current automated data science and AutoML solutions. We discuss how leveraging basic semantic reasoning on data in combination with novel tools for data science automation and foundation models can help with consistent and explainable data augmentation and transformation. We will describe how we are extending AutoML techniques to decision optimization, including sequential decision-making models and incorporating hard-to-model constraints and objectives. And finally, how semantics can help with challenges related to trust, bias, and explainabilty of both predictive and decision-making models.