Ensuring the authenticity of Extra Virgin olive oils is crucial due to the high risk of fraudulent practices associated with this valuable product. Traditional methods rely on physicochemical and organoleptic analyses, which are costly, time-consuming, and require specialized personnel. This study introduces probabilistic classification models utilizing Bayesian methods to enhance the reliability of Near Infrared Spectroscopy (NIRS) for olive oil (OO) quality control. Unlike traditional models, these methos allow the quantification of uncertainty, thereby improving decision-making precision in industrial applications. A total of 259 olive oils (104 extra virgin (EV), 71 virgin (V) and 84 lampante (L)) were analysed by two instruments with different optical configurations and sample presentation methods. Partial Least Square-Discriminant Analysis (PLS-DA) was applied to develop a two-step classification strategy: first, to discriminate non-LOO versus LOO categories, and then to predict the category of non-LOO samples (discriminating EVOO versus VOO). The models achieved a correct classification rate (CCR) of up to 86.36% for discriminating EVOO vs. VOO with the bench-top instrument, with more than half of the samples classified into their respective categories with a probability exceeding 75%, which highlights their effectiveness in ensuring the quality and authenticity of VOOs while optimizing resources in the olive oil industry. Similar results (81.82 %) were obtained for the portable device, despite differences in operational range, optical quality and price. The results demonstrate that probabilistic classification models can significantly improve the classification process by quantifying uncertainty, thereby complementing traditional methods and providing a robust framework for classifying olive oils categories.