Why does a model trained on principal components perform significantly worse than a model trained on true predictors?

Question

Machine learning

For a certain set of predictors, after the analysis of the main components, new predictors were obtained that describe approximately 99% of the variations in the original data. The model was trained (I don’t know if this is important, but it was a regression decision tree) on the original predictors and on the main components. The quality of the model on the principal components turned out to be about 17% worse than on the original predictors. What is the reason, because the main components describe almost the whole picture in the data?