The increasing sophistication of financial fraud schemes and the limitations of traditional rule-based detection methods necessitate the development of more adaptive and intelligent approaches for transaction analysis. Current detection systems struggle with evolving fraudulent patterns and generate false positives due to overly broad static rules. This paper proposes an autoencoder-based model for detecting suspicious transactions trained on non-fraudulent transaction data. The approach integrates both numerical transaction characteristics and textual descriptions as features, with textual data transformed into semantic embeddings using the Sentence-BERT model. Data preprocessing includes normalization and feature integration into a unified vector space, while the anomaly threshold is determined by maximizing the F1-score on validation data through cross-validation. Experiments conducted on a real-world dataset of 1,739 financial transactions from a retail enterprise, containing 104 fraudulent cases, demonstrate that combining numerical and textual features significantly improves the model’s ability to distinguish between normal and anomalous transactions compared to using individual data types. The model achieved high precision and recall rates, with fraudulent transactions showing distinctly higher reconstruction errors than non-fraudulent ones. The results confirm the practical applicability of the unsupervised learning approach and demonstrate its potential for implementation in auditing firms and financial institutions requiring automated transaction monitoring systems.