Recent explainable artificial intelligence work have often revealed the input features to which a model is attentive; however, lists of pixel or one-hot importance remain obscure to nonexperts and even domain specialists. This study proposes concept-approximate inverse model explanations (concept-AIME), an explanation framework that replaces raw features with a compact set of human-defined concepts (e.g., highly educated, managerial, and married). Concept-AIME operates in two stages. First, it learns a concept activation vector for each concept from a handful of positive samples and intermediate network activations and then develops a single approximate inverse operator that maps concept scores back to input features. Second, it constructs an analogous inverse operator for the target output (such as high-income probability) and, with one matrix product, instantly yields the percentage contribution of each concept to any individual prediction. In contrast to LIME or SHAP, which sample perturbations for every test point and stay bound to feature-level explanations, concept-AIME derives its inverse operator once, requires no gradients or internal parameters, and is therefore model-agnostic and sub-millisecond per instance after preprocessing. The method can immediately generate narratives such as “the ‘managerial’ concept raised the high-income score by 40%, whereas ‘married’ added 30%; lower education slightly suppressed the prediction.” By unifying global and local reasoning in a single linear formulation, concept-AIME offers a novel general-purpose technique that quantifies black-box decisions through user-chosen, semantically meaningful concepts, providing actionable, domain-relevant explanations for high-stakes applications in healthcare, finance, and public policy, where transparency is essential.