Existing Automated Service Composition (ASC) approaches typically require inputs to be in a designated form. These, namely tuples, pose challenges due to the significant divergence from the most commonly used and straightforward formats for expressing software requirements. In our previous work, we developed a rule-based approach that necessitated substantial resources for analyzing the content of requirements and establishing appropriate rules. Given the recent successes in field research involving large language models (LLMs)-where significant achievements have been made in real-time automatic text generation tasks-we propose leveraging LLMs for ASC to extract critical tuple-based information. We have created a new dataset to simulate everyday service demands and have established clear guidelines regarding service demand types (e.g., input and output). Moreover, we have implemented an appropriate workflow that optimizes LLMs performance. Our experiments and results demonstrate that our proposed LLMs-based approach not only achieves extraordinary performance and reliability at a lower cost but also outperforms the complex rule-based solutions that were previously employed.