PCA is a great tool to create linearly uncorrelated features. However, since it transforms the original features in other ones, you actually lose the business information. Sometimes you’d better keep a bit of collinearity in order to have some business-significant features that can be explained to a manager. Moreover, collinearity is not a real problem for some models like Random Forests. That’s why I prefer not to use PCA unless any other method doesn’t give any useful results.