With the increasing number of publications embracing the potential and broad applicability of Large Language Models (LLMs), the critical assessment of their limitations and true capabilities often remains underrepresented. Some studies even predict the shift towards LLM-based solutions over traditional Machine Learning (ML) approaches, arguing that LLM integration reduces the need for additional preprocessing and extensive training. However, some critical limitations of LLMs, which challenge their ability to substitute conventional ML, seem to be overlooked. One of these limitations is their limited ability to process numeric data and perform meaningful data analytics effectively. This paper investigates the challenges of applying LLMs directly to numerical data. The business objective of the investigated use case is to detect anomalous behavior in a manufacturing plant using production data. We begin with a literature analysis to assess the current research landscape, followed by empirical validation of the outlined research questions. Our findings indicate that the investigated LLMs struggle to process and analyze numerical data effectively, therefore limiting their applicability for specific use cases. Furthermore, we highlight the need for more rigorous evaluations of LLM capabilities in scientific research, specifically on unseen, real-world datasets, to provide a more objective assessment of their potential.