This study presents and assesses a Local Large Language Model combined with Retrieval-Augmented Generation (RAG) for automating the drafting of official travel authorization documents in a university administration setting. The goal is to assess if knowledge-based generation enhances regulatory compliance and numerical accuracy over standalone LLM drafting. A design science approach was used. Travel regulations and disbursement manuals were indexed in a vector database and integrated with a workflow layer. We evaluated the system with 30 real-world travel scenarios across three configurations: baseline LLM, workflow-enhanced LLM, and strict RAG-enabled LLM. Performance was assessed through format compliance rate, allowance calculation accuracy, regulatory reference accuracy, hallucination rate, and average drafting time. Proportion-based accuracy estimation and comparative performance assessment were used in quantitative analysis. Results indicate that strict RAG grounding reached 100% accuracy in allowance calculations, nearly perfect regulatory alignment, and a significant decrease in hallucination compared to baseline setups. Retrieval added slight latency, but the performance improvements in compliance-sensitive drafting were substantial. This study shows that local deployment based on knowledge is crucial for reliable administrative AI systems. Limitations involve assessment within one institution and reliance on the thoroughness of indexed regulations.