An efficient and accurate evaluation of feature importance method for artificial intelligent and machine learning models without high computational costs is a key research question of this work. Many existing Shapley-value-based explainable artificial intelligence (XAI) methods, while theoretically sound, require evaluating all \(2^{n}\) feature subsets, making them impractical for high-dimensional data. Additionally, most approaches typically treat feature importance and feature selection as separate tasks, and they often return a single optimal subset of features. The challenge is to design a model-agnostic framework that can deliver Shapley value-competitive importance rankings while simultaneously performing efficient feature selection and uncovering a range of near-optimal subsets of features. To address this challenge, we propose the Markovian Process Explainer (MPE), a novel XAI method grounded in the Markovian coalition process value. Our methodology models feature selection as a stochastic coalition-formation process, where features are sequentially added to or removed from a subset until an absorbing state-corresponding to an optimal or near-optimal feature subset and its associated model is reached. We empirically evaluate the MPE using forward and backward Key Performance Index (KPI) metrics, comparing its performance with established Shapley-value-based XAI methods. The results demonstrate that the MPE produces feature importance rankings that are consistently competitive with, and frequently superior to, those of existing methods. Moreover, the framework performs feature selection as an intrinsic byproduct of its process, rather than as a separate step. This work holds significant practical and theoretical importance. By reducing computational cost of the exact Shapley value computation, the MPE offers a scalable, model-agnostic solution that can be used in applications, from healthcare to finance, without sacrificing explanatory quality. The set of near-optimal subsets of features obtained as byproduct of the MPE enhances robust decision-making and strengthen trust in complex machine-learning and AI models.