To analyze customer behavior, handling the quantities of complex data exchanges, and taking into account the diversity of Call Detail Records (CDRs) lengths in a telecommunications environment, this paper proposes an innovative approach involving a machine-learning model SARIMAX (Seasonal Autoregressive Integrated Moving-Average with eXogenous factors) to perform predictive analysis of these calls with the aim of improving operational efficiency. The process of prediction model involves exploratory data analysis (EDA) to explore the dataset and identify anomalies, as well as autocorrelation (ACF) and partial autocorrelation (PACF) functions are applied in order to find the structure of the time series. Furthermore, we evaluate the performance of the model with mean absolute error (MAE), coefficient of determination (R2) and mean squared error (MSE). Finally, we use the Difference (DIFF), Square Root (SQRT) and the Cube Root (CBRT) transformations in order to turn these complex data into a smoother and more appealing idea that will improve the quality of our model. Its performance was promising as the proposed model revealed the least prediction errors with a good trade of between variance explained (R2) and minimizing errors (MAE, MSE). The outcomes validate the practical applicability and efficacy of our model to predict call duration and network capacity required.