Steel tubes produced in steelmaking plants are generally subjected to severe in-service conditions. Hence, quality control plays a key role in this process. The bottleneck is that this information is made available only after tube production from laboratory analysis. Given process complexity and current data availability, this work employs a series of machine learning techniques, namely neural networks, random forests and gradient boosting trees, to predict critical mechanical properties for steel tubes, namely yield strength, ultimate tensile strength and hardness. The model performance was kept high by combining different variable selection procedures. The prediction error was less than the inherent variability of each mechanical property, i.e., it is equal to 20 MPa for yield strength and ultimate tensile strength, and to 2 HRC, for hardness. This information in advance allows interventions before complete tube production contributing to more stable operations and, ultimately, to reduce rework and customer lead time. In sequence, an optimization problem for set point definition is illustrated. The neural predictive model previously identified for the yield strength was used in this application, exploring its predictive capabilities. The optimal solution yielded to lower amount of molybdenum and tube exit temperature from the tempering furnace, while keeping quality aspects, which means reduction in material and energy costs. Concluding, steelmaking processes, which are complex by nature, can strongly benefit from data-driven approaches, since data availability and computational processing are no longer a problem.