The use of modeling tools for integrated water resources management is a complex task due to the large number of processes involved in a basin. Moreover, these modeling tools commonly require information that is not readily available, such as illegal water withdrawals, or other data difficult to obtain, which results in groundwater models that fail to capture the aquifer dynamics. In recent years, machine learning algorithms have shown outstanding performance as prediction tools. Despite being questioned for not having a physical basis, they have been used in areas such as hydrology and hydrogeology (e.g., for flow prediction, rain forecast). Thus, the objective of this research is to estimate groundwater withdrawals using machine learning algorithms and integrated water management models. To achieve this objective, ensembles of groundwater levels were generated with a previously calibrated groundwater/surface water integrated model. Then, these ensembles were used as input parameters for Gaussian process regression (GPR) and artificial neural network (ANN) models to construct time series of water withdrawals throughout a basin. This method was applied in the Petorca and La Ligua basins, in central Chile, as they exhibit a contrasting reality in terms of water availability even when they have geographical proximity. Also, these basins are within an effective extraction monitoring program lead by the Chilean water authority that can be used to validate the users’ water withdrawal. Our results show that the GPR model, compared to ANNs, adequately estimates the spatiotemporal distribution of groundwater withdrawals in the pilot basins. Thus, the use of machine learning algorithms improves the performance of integrated water resources management models.