In this work, we leverage Machine Learning to accurately predict the 28-day compressive strength of co-grinded limestone-calcined-clay cement (LC3) in a production environment. We illustrate how ad-hoc measurements of chemical composition and particle size, including x-ray fluorescence, sieve analyses, Blaine and water demand, can be used for 28-day strength prediction. Our model reduces the prediction error by 30% compared to a baseline without Machine Learning. To obtain actionable insights and identify optimization potential, we apply techniques of eXplainable Artificial Intelligence and model interpretability. For the evaluated cement plant dataset, the most important features for strength prediction include LOI and various proxies for particle sizes. A cautious interpretation of these correlations suggests variable limestone content as well as over- and under-grinding as the main sources of strength variability, in contrast to a relatively well-optimized clinker. Overall, this case study demonstrates how interpretable Machine Learning can give key insights for monitoring and optimization of LC3 quality in production.