In this paper, we introduce a relevant approach for payroll fraud detection, leveraging Large Language Models (LLMs) to generate synthetic fraud data scenarios that are not text-based. For the first time, LLMs are used to create structured data that present real payroll patterns without relying on textual tasks. These synthetic data integrate seamlessly with real non-fraud data. In fact, this is achieved through an extensive data analysis phase that captures natural feature correlations and patterns, ensuring that the generated data mirror realistic data. The synthetically generated data are then used to fine-tune a finance pre-trained LLM, enabling the generation of scalable fraud schemes. Therefore, they are combined with historical non-fraud data, and then input into a meta multilabel classifier model to categorize fraud scenarios. Obtained results are evaluated against those from a standard multi-label classifier model. The comparison confirms that our approach has higher performance compared to a decision tree multi-label classifier on a real-world test dataset across nearly all fraud labels. Indeed, the proposed approach can easily generalize to other fraud domains and could be extended to unlabeled data, ensuring both data fidelity and diversity.