In the retail sector, inventory audits are of paramount importance as they exert a direct or indirect influence on various dimensions of operational efficiency. Therefore, since the inception of the industry, one of the main objectives has been to ensure inventory accuracy. In line with this objective, inventory counts have become the most common method for reducing financial losses, eliminating inefficiencies in the process, and achieving the desired level of inventory accuracy. However, the occurrence of fraudulent activities during these counts has the potential to compromise the reliability of the data, thereby diminishing the anticipated benefits of inventory counts. This study proposes a machine learning-based anomaly detection model for the retail industry, with the objective of enhancing the reliability of inventory counts and reducing the labor required to audit the counts. The study utilizes a dataset comprising historical inventory counts from 3,000 stores in Turkey, along with transaction records that resulted in inventory changes at these stores. We analyze the dataset to identify anomalous patterns, extract relevant features, and apply unsupervised learning models such as Isolation Forest, Empirical Cumulative Distribution-based Outlier Detection (ECOD), and Local Outlier Factor (LOF) for anomaly detection. With the developed anomaly detection models, we are able to detect fraudulent stock counts with high accuracy, thereby significantly reducing the necessity for manual audits. Following a year of active implementation, we observe a 1.16% decrease in fraudulent inventory counts and a 0.13% increase in inventory accuracy.