<p>Accurate prediction of concrete properties is a critical aspect of structural engineering. Conventional neural network models, while effective in various predictive tasks, often fall short in this context because they do not distinguish between the risks associated with underestimation and overestimation of strength. Existing codes of practice in civil engineering advocate a conservative estimation approach, prioritizing underestimation to preserve safety margins. Standard neural network models, which treat all prediction errors uniformly, are therefore not aligned with this conservative bias, which motivates the present study. The aim of this research is to design a custom loss function that embeds domain-specific safety requirements into artificial intelligence models. This function penalizes underestimations less severely than overestimations, reflecting the conservative principles of structural design. The problem formulation emphasizes the importance of conservative concrete strength prediction within the broader context of structural safety. Rubberized concrete is adopted as a case study to investigate the performance of the proposed custom loss function and to benchmark it against traditional loss functions. The study results show that the developed custom loss function effectively addresses conservative estimation needs by reducing the overestimation ratio from about 25–100% with traditional approaches to as low as 1%. At the same time, prediction accuracy remains comparable to conventional loss functions. Accordingly, the proposed custom loss function can serve as a technical enhancement that improves prediction behavior in line with conservative design philosophy and supports the reliable use of neural network models as one component within the structural design process.</p>