This paper explores the application of large language models (LLMs), particularly GPT-4, in the context of Thai land and building tax computation. We address challenges arising from the complexity of Thai tax regulations, the scarcity of Thai-language training data, and the structured, multi-variable nature of tax computation tasks. To enhance model performance, we employ few-shot prompting, decomposition-based reasoning, and retrieval-augmented generation (RAG). We further introduce a JSON-style output format to support structured conversational interactions, allowing the model to identify missing user inputs and guide data collection. A secondary o3-mini pass is used as an answer verifier, helping to assess the logical and numerical correctness of each output. Experiments conducted on an augmented dataset show that the combination of decomposition, structured output formatting, and LLM-based verification leads to substantial improvements in explanation quality and overall computation accuracy, confirming that the integration of these techniques not only enhanced explanation clarity but also ensured greater reliability and correctness in tax computation results.