We study the problem of predicting in real-time customers' waiting time. We focus on (1) service systems where customers are scheduled for service ahead of time and where real-time and scheduling data is available to the model (e.g., in cancer care); and (2) the predictions are intervals that are dynamically updated over time (a format supported by practice and research). We collaborate with a large cancer hospital and conduct an extensive set of numerical experiments using recent operational data. We compare a broad set of methodologies and report on their performance in terms of interval width and accuracy, two contrasting yet desirable objectives. We further develop and evaluate a prototype in a preliminary field experiment.&nbsp; <div> Our study shows that Random Survival Forest outperforms other Machine Learning models, establishing itself as state-of-the-art. We find that despite the availability of a rich dataset, creating accurate predictive models is challenging due to stochasticity, heterogeneity, the complex service process, and limitations of the available data. A simulation study suggests that the quality of prediction could not be improved due to an irreducible error. Our experiments also show that while real time data substantially improves the prediction quality, hospitals can still benefit from such models without investing in infrastructure that supports real time tracking of patients and providers. Finally, we discuss lessons learned from the development process, including the customization of predictive models to improve performance and usability, the use of Explainable AI, and ways in which future endeavors could potentially create better models. </div>