TnT-LLM is a two-phase framework that uses large language models (LLMs) to automate the generation and assignment of labels for text mining tasks. The first phase involves zero-shot, multi-stage reasoning to iteratively produce and refine a label taxonomy. The second phase uses LLMs as data labelers to generate training samples for lightweight classifiers. TnT-LLM was applied to analyze user intent and conversational domain in Bing Copilot, an open-domain chat-based search engine. Extensive experiments showed that TnT-LLM generates more accurate and relevant label taxonomies compared to state-of-the-art baselines and achieves a favorable balance between accuracy and efficiency for classification at scale. The framework is adaptable and modular, requiring minimal human intervention. The paper also discusses the challenges and opportunities of using LLMs for large-scale text mining in real-world applications. The framework was evaluated using deterministic automatic metrics, human evaluation metrics, and LLM-based evaluations. Results showed that TnT-LLM outperformed other methods in label accuracy and relevance. The lightweight classifiers trained on LLM annotations achieved comparable or better performance than directly using LLMs as classifiers, with higher scalability and model transparency. The paper also discusses the use of LLMs as annotators and evaluators, highlighting their effectiveness in certain tasks but limitations in others. The framework has the potential to significantly impact AI research and applications in text mining by enabling efficient and scalable label generation and classification.TnT-LLM is a two-phase framework that uses large language models (LLMs) to automate the generation and assignment of labels for text mining tasks. The first phase involves zero-shot, multi-stage reasoning to iteratively produce and refine a label taxonomy. The second phase uses LLMs as data labelers to generate training samples for lightweight classifiers. TnT-LLM was applied to analyze user intent and conversational domain in Bing Copilot, an open-domain chat-based search engine. Extensive experiments showed that TnT-LLM generates more accurate and relevant label taxonomies compared to state-of-the-art baselines and achieves a favorable balance between accuracy and efficiency for classification at scale. The framework is adaptable and modular, requiring minimal human intervention. The paper also discusses the challenges and opportunities of using LLMs for large-scale text mining in real-world applications. The framework was evaluated using deterministic automatic metrics, human evaluation metrics, and LLM-based evaluations. Results showed that TnT-LLM outperformed other methods in label accuracy and relevance. The lightweight classifiers trained on LLM annotations achieved comparable or better performance than directly using LLMs as classifiers, with higher scalability and model transparency. The paper also discusses the use of LLMs as annotators and evaluators, highlighting their effectiveness in certain tasks but limitations in others. The framework has the potential to significantly impact AI research and applications in text mining by enabling efficient and scalable label generation and classification.