This article explores the challenges of few-shot prompting in language models, specifically in customer service intent classification tasks, and how to improve model performance by addressing noisy and erroneous examples. The authors use the Davinci Large Language Model from OpenAI to classify the intent of customer service requests at a large bank, but encounter issues with the accuracy of their LLM predictions due to real-world data being messy and error-prone. They find that using data-centric AI algorithms via Cleanlab Studio to ensure only high-quality few-shot examples are selected for inclusion in the prompt template significantly boosts model performance. The authors demonstrate that modifying the prompt or removing examples alone cannot guarantee optimal model performance, but instead, data-centric AI tools like Cleanlab Studio can identify and correct label issues, resulting in improved accuracy.