At RSI, there are a variety of applications of outlier detection (or anomaly detection). In our analytics work, these applications include: fraud scoring, collections analysis, audit selection, and others. In the audit selection framework, given the right feature data, machine learning (ML) can be a powerful tool for assisting in the selection of potentially productive audits (which at least pay for the auditor's time). However, training an accurate supervised ML model relies on having sufficient number of labeled examples in the dataset. In the case of audits, a single labeled example can take an auditor anywhere from a few days to a few weeks to provide the change status (underpaid, no change, or overpaid) and the amount due. This audit result can be appealed, etc.; and so the time involved to obtain this labeled example can become inordinate. Thus, even with ten years of data, the number of labeled examples in the dataset may be insufficient for a typical supervised ML model training. Thus, we need to use features of the unlabeled dataset to provide more information to the supervised model training process. This can be done using clustering techniques to find general "behavioral groups", as well as outliers. These unlabeled data "derived features" are incorporated as a part of the supervised audit model. In a particular effort, the RSI analytics team found that including such clustering and anomaly features in the supervised process improved our accuracy in classifying a potential productive audit from around 65% to about 95%. At RSI, we continue to apply unsupervised approaches to other application areas where labeled data is spare, but where there is still structural information in the unlabeled data, like clusters and anomalies. We are also applying an outlier/anomaly detection as a secondary model to assist in identifying fraud. In this case, there is sufficient data, but as fraudulent behaviors change, the supervised models will suffer from performance drift. Our anomaly detection model assists in identifying fraud outside of the behaviors seen during supervised training, as well as being an indicator that the primary supervised model needs to be tune and/or include newer data in the training process.