We build models because we want to highlight the most relevant factors of the circumstances. In a broad sense, language, music, painting are all different forms of modeling: they stress certain aspects of reality by ignoring the rest.
Models must be useful. Unlike simplicity and generality, usefulness is trickier to assess. The challenge arises from the parallel of two worlds: the real and the model world; so there is a gap one must be able to cross.
The main idea of modeling is that, if we isolate the substantive factors of a system into a model, then the outcomes generated by the model should also find counterparts in the real world. Or, if we observes certain outcomes in the real world, and the model produces similar outcomes, then these substantive factors should be determinants in the real world, too.
Both arguments rely on inductive inference. In general, the model and real world should share similarities in structure, dynamics, and outcomes. Yet these similarities are no guarantee that two worlds move in lock step. Cases abound of otherwise. Thus, relevance has to be taken by faith, not by the logic reasoning. There are gaps that one must be willing to cross.