The construction industry frequently suffers cost and schedule overruns. Contributing factors include fragmented risk processes and the laborious manual review of large volumes of unstructured text. In 2022, the DREAM model demonstrated that combining machine-learning classifiers with expert knowledge can uncover the root causes of delays more systematically than manual review. However, it required extensive retraining for each new context and was prone to bias from experts. Recent advances in large language models, particularly GPT-4, offer zero or few-shot adaptability across vast unstructured text, eliminating the need for bespoke annotation. This paper evaluates whether an off-the-shelf GPT-4 model, used in zero-shot mode, can accurately classify causes of delay (CoDs) from meeting minutes without fine-tuning. A protocol for prompt engineering with iterative self-consistency checks is implemented to maximize reliability. Experimental results, including confusion-matrix analysis, demonstrate that ChatGPT-4 outperforms DREAM by closing the performance gap while significantly reducing annotation effort. By quantifying annotation savings and characterizing residual biases, we demonstrate that modern large language models (LLMs) enable scalable, data-efficient AI support in construction risk management.