In intent classification tasks with constrained data, data augmentation can emerge as a powerful tool for enhancing Machine Learning (ML) model performance. In this context, existing augmentation methods using off-the-shelf general-purpose models can be improved. This paper presents a new utterance data augmentation technique that uses off-the-shelf general-purpose large language models (LLMs) in a two-step manner. First, paraphrases are obtained by using an LLM prompt, generating semantically similar utterances with variations in wording and syntax. Second, synonym replacement is performed using another LLM prompt, further augmenting specific target words. Experiments on two financial trading compliance intent classification problems demonstrate that this proposal is an improvement over previous approaches.