The purpose of the study is to develop an approach to constructing multi-modal machine learning models for assessing soft skills of business simulation players using game logs. The study was performed by analyzing the logs of business simulation “Corporate Governance”, which simulates the management of an enterprise in a real market. Within the framework of the study, business simulation is considered not as a learning game that forms competencies, but as a diagnostic one for assessing the players’ soft skills. The approach allows taking into account simultaneously each player’s individual strategy and the overall team scores in the assessment. An approach to the application of machine learning methods for analyzing business simulation logs is proposed, based on constructing a meta-algorithm that takes into account various types of input data. Individual player actions data are considered as action sequences and are treated by text data processing methods. As research implications, this paper presents a new integrated conceptual approach, which can be useful for studies focused on recruitment techniques and employee skills diagnostics. Currently, the player’s competencies are actually measured manually, rather than using tools for automated assessment. It is time-consuming and costly, especially when it is necessary to conduct mass business simulations. This research provides guidance for automating the process of assessing player skills thus delivering benefits of practical importance.