Framework for integrating industry knowledge into a large language model to assist construction cost estimation

The construction industry generates a large amount of data across projects produced by digital devices, tools, and methods, and this volume is rapidly increasing. However, the industry lags behind in adopting data-driven technologies. On the other hand, the rapid advancement of generative AI (GenAI) in recent years, especially state-of-the-art large language models (LLMs), shows great potential and has been increasingly adopted in many industries; however, the construction industry is behind in adoption. While academic studies have proposed various machine learning applications for construction, industry implementation has lagged due to a disconnect between these proof-of-concept developments and practical industry needs. Also, to effectively leverage advanced LLMs in construction workflow, the gap remains in assessing the current implementation status-maturity level, industry’s interest, or priorities for such technologies. Among critical workflows, cost estimation, a core function that carries high financial decision control and risk, stands out as a high-priority area where inefficiencies, repetitive processes, and intuition-driven decisions continue to hinder productivity and accuracy. To address this gap, this dissertation works on two guiding questions: What is the current status of the implementation of data-driven technologies in industry? And how can we integrate construction industry knowledge into a large language model to advance industry practice? This dissertation addresses that gap by developing a framework to integrate construction industry knowledge, specifically a workflow, into an LLM for assisting with cost estimation tasks, using generative AI as a scalable, human-in-the-loop solution. The research follows a mixed-methods approach. First, a convergent mixed-methods study assessed the current status of data science adoption in the construction industry. Results revealed a low level of implementation despite high interest, with cost estimation and scheduling identified as the most critical domains for AI integration. Building on this insight, the second phase mapped existing cost estimation workflows and, through interviews with industry subject matter experts (SMEs), identified recurring burdens in the industry estimation workflow. To address these challenges, the dissertation developed a GenAI-assisted estimation framework structured around three major estimation stages: conceptual estimation, subcontractor evaluation, and construction-phase cost updates. It tested whether current state-of-the-art general-purpose LLMs could follow real-world estimation tasks under a zero-shot setting. Results showed limitations in completeness, accuracy, and following the workflow in sequence. This limitation led to the development of a modular chain-of-thought (CoT) prompting approach that breaks complex estimating tasks into smaller, sequential reasoning steps. This improved the performance of LLM significantly, increasing human evaluation confidence scores and improving results across multiple language model evaluation metrics. Building on this success, the final phase developed and validated a customized AI assistant- CNST-GPT- for cost estimation workflow using the GPT-4o model and refined through subject matter experts’ feedback. Validation through industry workshops showed statistically significant reductions in estimator workload across all identified burdens, with strong effect sizes on time, mental effort, and psychological stress. Feedback from professionals also guided refinements to improve output consistency, reduce hallucinations, and ensure alignment with firm-specific estimation practices. This dissertation makes five key contributions: (1) an empirical assessment of data science implementation in construction, (2) a mapped model of current estimating workflows and pain points, (3) a GenAI-integrated estimation framework, (4) a modular CoT prompting prototype for domain-specific LLM tasks, and (5) a validated prototype AI assistant, CNST-GPT. Together, these contributions advance both theor and practice by bridging the implementation gap between generative AI and construction industry estimating practice. This dissertation demonstrates how LLMs, when guided with instructions, domain knowledge, and designed for human collaboration, can effectively execute critical construction management functions. Overall, this early study serves as foundational literature to encourage subsequent research expansion in LLM applications in other workflows within the construction industry and its allied architecture and engineering domains.