An important factor in a circular agricultural system is the efficient use of animal manure. Until now, the applied quantity of manure is regulated by law at farm level, based on fixed phosphorus (P) application norms. However, a first step towards more efficient manure application is to better balance P input and output at field level by predicting future P yields. Machine learning techniques can be useful in this respect, because they can be trained with many variables without prior knowledge regarding their interrelationship. This study’s objective, therefore, was to predict P yields based on detailed records of on-farm data as recorded on an experimental farm combined with open source weather data. The dataset contained 657 records of annual crop yields per field between 1993 and 2016, and the boosted regression model was used for model development. Validation on the final five years of the dataset resulted in an RMSE of 7.3 kg P per ha per year, an R-squared of 0.46 and a correlation between observed and predicted values of 0.68, outperforming legal norms. We conclude that with the limited but detailed data available, prediction of P yield, and therewith, defining flexible P application norms before first manure application, is already feasible. This conclusion, together with the expected increasing availability of data through proximal and remote sensing technologies, opens the way to further improve nutrient management and move towards circular agriculture in the future.