Potential of Large Language Models for Construction Quotation Workflows

Abstract: Large language models (LLMs) can automate text-intensive tasks in construction. However, preparing investigation quotations remains manual and time-consuming, with outputs varying between engineers due to non-deterministic steps like item selection and text drafting. This study proposes an LLM-assisted human-in-the-loop (HITL) workflow and evaluates several models (GPT 4.1, GPT 5, Claude Sonnet 4.5, and a fine-tuned open-source model, QwQ 32B) on 596 inquiry-quotation pairs. GPT 5 achieved the highest item-selection F1 score, while Claude 4.5 delivered comparable accuracy with the fastest processing time. This approach improves quotation efficiency and consistency, demonstrating practical benefits for construction workflows.