Machine Learning models can effectively be used in the public sector to classify user-uploaded documents for more efficient administration. However, the nature of user-uploaded data is non-stationary, as the input data stream may be affected by external influences, ranging from redesigns of official documents to geopolitical events that impact the user demographics interacting with the system. When the incoming data deviates from the data used at training, it may introduce concept drift making the models performance degrade over time. Effectively detecting concept drift can be the first step in a model adaptation strategy, as it indicates that a model update is needed.This thesis investigates practical methods for concept drift detection in a document classification domain, focusing on the feasibility to use other observations variables than prediction accuracy which is traditionally used, such as predicted labels and confidence levels. A pilot experiment was conducted using the Fashion MNIST dataset in order to validate the experimental setup and the drift detectors ADWIN, KSWIN and Page Hinkley, before performing analogous experiments using data from a real document classification application.The findings suggest that while confidence levels and predicted labels can be used to detect concept drift, ADWIN may not be the best detec- tor for these observation variables with the parameter values explored. KSWIN and Page Hinkley showed potential but also produced high false positive rates. Differences in the results between the Fashion MNIST experiment and the document classification experiment were observed and underscore the importance of tuning and validating a drift detec- tor for its intended environment. The findings highlight the potential to use confidence levels and predicted labels, emphasizing the need for robust parameter tuning and domain-specific knowledge in developing effective drift detection strategies in real world applications.