Adaptive neuro-fuzzy inference systems (ANFIS) can efficiently predict adsorption behavior, reducing the need for extensive experimental trials. In this study, copper (II) adsorption data collected at various pH values, adsorbent doses, contact times, and initial concentrations were analysed via MATLAB. The membership functions and optimal number of member functions were evaluated to enhance the Box–Behnken experimental design (BBD) and generate accurate ANFIS-based adsorption predictions with minimal error. For each input variable, the BBD matrix was assessed across all possible combinations to determine the optimal number of member functions. Regression models were developed for each member function, incorporating both the root mean square error (RMSE) and the number of member functions per input. Four common membership functions, namely, triangular, trapezoidal, Gaussian, and Gaussian 2, were investigated. Analysis of variance (ANOVA) confirmed that models using triangular membership functions were statistically significant at the 95% confidence level. Predictions via ANFIS were optimised with a triangular membership function configuration of 12–12–12, which resulted in the maximum copper (II) adsorption capacity. Furthermore, the relationship between member function numbers and RMSE values was visualised via three-dimensional plots for the triangular membership function, providing deeper insight into adsorption performance.