According to CCSK, what is the recommended approach for classifying data before moving it to the cloud?