Millions of people are exposed to the potential risks of high arsenic (As) concentration through groundwater intake. Numerical models are powerful supporting tools for decision makers to make informed decisions that could help to minimize such risk, yet the ubiquitous presence of heterogeneity creates uncertainty in the model predictions. This is demonstrated through this presentation, which summarizes the key results of a modeling analysis using 3D multicomponent reactive transport of the shallow aquifer of Venetian Alluvial Plain (VAP). The VAP is notoriously affected by As contamination, characterized by a patchy distribution with variable extensions and concentrations, sometimes exceeding the WHO recommended limits for drinking water. Within the VAP, we focused in detail on an agricultural zone nearby the Venice lagoon, affected by As contamination (called “Aree Agricole West”, AAW). The available data, collected by several hydrogeological surveys, show a spatial and temporal variability of As concentration, which can be associated to a variety of hydro-geochemical processes such as redox variations, sorption or reductive dissolution of As-rich iron oxy-hydroxides. The 3D reactive transport model showed a strong dependence between the aforementioned processes and subsurface heterogeneity. The material distribution, indeed, plays an import role affecting the arising of the main chemical reactions. There is a clear spatial and temporal distribution of dissolved arsenic as a function of the subsurface heterogeneity. The results are explained considering strong effects of oxy-reductive potential on arsenic mobility, likely correlated to organic matter degradation. Depending on the recharge seasonality, the model suggests the uprising of reduced condition, as affected by other mechanisms such as reductive dissolution of iron hydroxides, ion exchange and sorption processes, causing arsenic mobilization. The model also suggests the importance of a detailed characterization of the site, which constrains input parameters that could otherwise create further uncertainties in the model outputs.