The current advances of environmental DNA (eDNA) bring profound changes to ecological monitoring and provide unique insights on the biological diversity of ecosystems. The very nature of eDNA data is challenging yet also revolutionizing how biological monitoring information is analysed. In particular, new metrics and approaches should take full advantage of the extent and detail of molecular data produced by genetic methods. In this perspective, machine learning algorithms are particularly promising as they can capture complex relationships between the multiple environmental pressures and the diversity of biological communities. We investigated the potential of a new generation of biomonitoring tools that implement machine-learning techniques to fully exploit eDNA datasets. We trained a machine learning model to discriminate between reference and impacted communities of freshwater macroinvertebrates and assessed its performances using a large eDNA dataset collected at 64 standard federal monitoring sites across Switzerland. We show that a model trained on eDNA is significantly better than a naive model and performs similarly to a model trained on traditional data. Our proof-of-concept shows that such a combination of eDNA and machine learning approaches has the potential to complement or even replace traditional environmental monitoring, and could be scaled along temporal or spatial dimensions.