Traditionally, agent-based modelling and simulation relied on using utility function in agents' decision-making process. Some drawbacks in this process are identified, and a potential remedy to the issue is proposed. This paper introduces a methodological framework for building a hybrid agent-based model that aims to overcome some of the elaborated problems related to the usage of a utility function. In the proposed approach, a machine learning algorithm substitutes the utility function, thus providing a possibility to use various algorithms. The proposed methodological framework has been applied to a case study of churn in a telecommunications company. Three models have been created and used for simulation experiments, two using the proposed methodology and one using utility function. The pros and cons of different approaches are identified and discussed.