Some of the most common and typical issues in the field of intelligent tutoring systems (ITS) are (i) the correct identification of learners’ difficulties in the learning process, (ii) the adaptation of content or presentation of the system according to the difficulties encountered, and (iii) the ability to adapt without initial data (cold-start). In some cases, the system tolerates modifications after the realization and assessment of competences. Other systems require complicated real-time adaptation since only a limited number of data can be captured. In that case, it must be analyzed properly and with a certain precision in order to obtain the appropriate adaptations. Generally, for the adaptation step, the ITS gathers common learners together and adapts their training similarly. Another type of adaptation is more personalized, but requires acquired or estimated information about each learner (previous grades, probability of success, etc.). Some of these parameters may be difficult to obtain, and others are imprecise and can lead to misleading adaptations. The adaptation using machine learning requires prior training with a lot of data. This article presents a model for the real time automatic adaptation of a predetermined session inside an ITS called AI-VT. This adaptation process is part of a case-based reasoning global model. The characteristics of the model proposed in this paper (i) require a limited number of data in order to generate a personalized adaptation, (ii) do not require training, (iii) are based on the correlation to complexity levels, and (iv) are able to adapt even at the cold-start stage. The proposed model is presented with two different configurations, deterministic and stochastic. The model has been tested with a database of 1000 learners, corresponding to different knowledge levels in three different scenarios. The results show the dynamic adaptation of the proposed model in both versions, with the adaptations obtained helping the system to evolve more rapidly and identify learner weaknesses in the different levels of complexity as well as the generation of pertinent recommendations in specific cases for each learner capacity.