The paper examines the model of energy cost management in a four-level control system of company, includes its boss with the advisor on top level, administrator on the middle level, and managing director of the plant at the bottom level. Neither the boss nor the advisor knows minimal administrator and plant stochastic energy costs, and need to learn to control them. The administrator knows minimal stochastic costs better than the boss and the advisor. So the administrator can manipulate costs in order to influence the results of learning the advisor and the boss in own favor. But the administrator itself does not know the minimum stochastic costs in plant. This can be used by its director to achieve own goal. So the administrator also needs to learn to control the managing director. All of the interests of such active elements are reflected in the model by introducing goal functions. On the basis of this model, sufficient conditions have been found for the synthesis of a hierarchy of mechanisms for managing energy costs, ensuring the use of stochastic possibilities of reducing these costs. In these conditions, the boss needs training with the help of a self-learning consultant. Through this training, the boss can rate the administrator's costs. This mechanism encourages the administrator, firstly, to minimize overhead energy costs and, secondly, to implement an adaptive cost savings mechanism on the plant. Such an adaptive mechanism includes standardization and stimulation procedures that encourage the managing director to minimize the energy costs of the plant. The application of this approach is illustrated by the example of managing energy costs in a wagon repair company.