Application of Large Language Models in the AECO Industry: Core Technologies, Application Scenarios, and Research Challenges

Abstract

As projects in the architecture, engineering, construction, and operations (AECO) industry grow in complexity and scale, there is an urgent need for more effective information management and intelligent decision-making. This study investigates the potential of large language models (LLMs) to address these challenges by systematically reviewing their core technologies, application scenarios, and integration approaches in AECO. Using a literature-based review methodology, this paper examines how LLMs—built on Transformer architecture and powered by deep learning and natural language processing—can process complex unstructured data and support a wide range of tasks, including contract analysis, construction scheduling, risk assessment, and operations and maintenance. This study finds that while LLMs offer substantial promise for enhancing productivity and automation in AECO workflows, several obstacles remain, such as data quality issues, computational demands, limited adaptability, integration barriers, and ethical concerns. The paper concludes that future research should focus on improving model efficiency, enabling multimodal data fusion, and enhancing compatibility with existing industry tools to realize the full potential of LLMs and support the digital transformation of the AECO sector.

1. Introduction

The architecture, engineering, construction, and operations (AECO) industry is a complex and fragmented sector, where each project phase involves multiple stakeholders, diverse tasks, and distinct information requirements. AECO projects generate a vast amount of documentation, including blueprints, specifications, contracts, change orders, and other project-related documents. These documents are often unstructured and exist in various formats, making their management and utilization a challenging task [1]. Throughout a building’s lifecycle—from design and construction to operation and maintenance—managing and utilizing unstructured text such as contractual clauses, technical descriptions, material reports, and site progress reports is essential. Effective information management is crucial for project success; however, due to the complexity and multi-stage nature of construction projects, this remains a significant challenge [2].

In the AECO industry, the increasing scale and complexity of construction projects have led to an explosion of project-related information, making it difficult for managers to track and process key data efficiently. Traditional natural language processing (NLP) techniques in AECO primarily rely on rule-based algorithms and statistical models [3]. These traditional NLP methods struggle with complex domain-specific content and require extensive manual intervention, especially in processing contractual clauses, technical documentation, and project communications. Recent advancements in artificial intelligence (AI), particularly in conversational AI (CAI) and generative AI (GAI), have paved the way for more scalable and intelligent information management approaches [4].

The emergence of large language models such as BERT and GPT has opened new possibilities for processing unstructured data in AECO [5]. Compared with traditional NLP approaches, LLMs leverage deep learning techniques to automatically learn intricate language patterns and contextual relationships from vast corpora, thereby demonstrating superior language comprehension and generation capabilities [6]. These models efficiently process complex construction-related textual data without heavy reliance on manually defined rules or domain-specific feature extraction. Through pretraining and fine-tuning, LLMs have shown remarkable performance in AECO applications, significantly enhancing information processing efficiency and accuracy—particularly in tasks such as unstructured text processing, information extraction, and intelligent decision support [7,8].

By leveraging deep learning techniques, these models can understand, generate, and analyze complex natural language, greatly facilitating automated text processing [9]. Moreover, LLMs support semantic understanding and context-aware reasoning, which are particularly useful in AECO domains where interdisciplinary collaboration and precision are essential [10]. In this domain, the application of LLMs extends beyond information extraction and text classification to tasks such as risk assessment, contributing to the enhancement of intelligent project management, reduction of human errors, and overall improvement of project efficiency [11].

This paper reviews the current state of LLM applications in the AECO industry and discusses its potential and challenges in information processing, project management, and intelligent decision-making; the roadmap can be seen in Figure 1. This study aims to:

(1). Review the development status of LLMs in the AECO industry, analyze their core technologies and typical application scenarios, and explore existing technical challenges and industry demands.

(2). Evaluate the practical value of LLMs in AECO, summarizing their advantages in information processing, intelligent decision-making, and project management while providing key technical references for real-world deployment.

(3). Explore the future development of LLMs in AECO, investigating technological innovations, system integration, and interdisciplinary collaborations to provide insights for future research and drive the intelligent transformation of the industry.

2. Materials and Methods

This study aims to explore the application of large language models (LLMs) in the AECO industry and their future development trends. In this study, the six-step literature review methodology proposed by Machi and McEvoy [12] was used to ensure a systematic and rigorous approach to reviewing existing literature, which has also been adopted in prior research, such as Onatayo’s study on BIM implementation efforts and strategies in selected countries [13] as well as their comprehensive review on the applications of generative AI in architecture, engineering, and construction, which discusses emerging trends, practical implications, educational considerations, and the necessity for upskilling in the industry [13]. (1) The scope of the literature review was delineated, with a specific emphasis on examining trends in the adoption of generative LLMs within the AECO industry, as well as their practical implications. (2) A comprehensive literature search was conducted using the Web of Science and Google Scholar databases. Multiple keyword combinations were employed to ensure extensive coverage of relevant academic publications. (3) A systematic quality assessment of the retrieved literature was performed. Diverse sources, including peer-reviewed journal articles, conference proceedings, and book chapters, were incorporated to ensure a rigorous and comprehensive data collection process. (4) The selected literature was subjected to an in-depth analytical review, with key insights extracted regarding prevailing trends and the impact of LLM applications within the AECO industry. (5) The findings underwent a thorough evaluation to ensure reliability and validity, employing established analytical frameworks and methodological rigor. (6) Finally, CiteSpace (version 6.1.6) was utilized for bibliometric and cluster analysis of the selected literature, generating structured insights that were critically examined to assess their potential implications for the AECO industry. Figure 2 presents the six-stage systematic review process of the literature.

The inclusion criteria encompassed peer-reviewed journal articles, conference proceedings, and book chapters published between 2018 and 2025, focusing on the application of LLMs within the AECO industry. Despite adhering to a rigorous selection process, we acknowledge certain limitations in our literature coverage. The reliance on English-language publications may introduce language bias, potentially overlooking significant research disseminated in other languages. Additionally, the concentration on studies indexed in databases like Web of Science and Google Scholar could result in geographic bias, favoring research from regions with higher representation in these databases. These biases may affect the generalizability of our findings across diverse global contexts. Future research endeavors should strive to incorporate multilingual sources and broader database searches to mitigate these limitations and provide a more comprehensive understanding of LLM applications in the AECO industry.

In the literature search phase, various keyword combinations were used to improve the efficiency of the literature review filtering process, reducing the manual screening required. The index term paper citation counts are as follows: (“large language model” OR “LLM” OR “GPT” OR “BERT”) AND (“construction engineering” OR “architecture” OR “civil engineering”). In the actual search process, the following results were obtained: “large language model” + “construction engineering” yielded 34 papers, “large language model” + “civil engineering” yielded 15 papers, “GPT” + “civil engineering” yielded 5 papers, “BERT” + “construction engineering” yielded 9 papers, and “BERT” + “civil engineering” yielded 12 papers. In total, 75 relevant papers were extracted and compiled.

Based on the classification framework presented in Figure 3, the subsequent sections are structured around four key thematic dimensions: (1) core technologies, (2) application scenarios, (3) data and information processing, and (4) toolchains and system integration.

This categorization reflects the major research within the existing works and also facilitates a more systematic and granular discussion of how large language models (LLMs) are being developed and deployed across the architecture, engineering, construction, and operation (AECO) domain. Each of the four dimensions corresponds to a critical pillar in the lifecycle-oriented digital transformation of the industry, from the foundational technological enablers, through domain-specific application contexts, to the supporting of data ecosystems and the integration of LLMs into broader software and operational environments.

3. LLMs in the AECO Industry: Core Technology, Application Scenarios, and System Integration

3.1. Core Technologies

3.1.1. From Traditional NLP to LLMs

In the field of architectural engineering, traditional natural language processing (NLP) techniques rely heavily on rule-based algorithms and statistical models [3]. With the advancement in AI, there has been significant improvement in the development of conversational AI (CAI) and generative AI (GAI). Conversational AI deals with the application of NLP to enable computers to understand and interact with humans in a conversational way using natural language. While these approaches have achieved some results in handling some scenario-specific textual tasks (e.g., lexical annotation, named entity recognition, information retrieval) [4], they also have obvious limitations. Rule-based methods require manual writing of a large number of rules, which makes it difficult to cope with complex and varied architectural text data, such as contracts, design specifications and project reports, etc. Furthermore, statistical methods (such as hidden Markov models and naive Bayes classifiers) [14] rely on predefined feature extraction techniques and struggle to capture deep semantic relationships and contextual dependencies in construction texts [15]. Facing the complexity of unstructured textual information, traditional NLP in the construction domain shows the limitations of low processing efficiency and poor generalization ability.

Large language models are advanced language models with massive parameter sizes and exceptional learning capabilities. The core module behind many LLMs such as GPT-3, InstructGPT, and GPT-4 is the self-attention module in Transformer [15]. Compared with traditional NLP, large language models can automatically learn complex language patterns and contextual relationships from large-scale corpora through deep learning. Large language models can automatically learn complex language patterns and contextual relationships from large-scale corpora through deep learning techniques, thereby possessing stronger language understanding and generation capabilities. These models are able to handle complex architectural textual data more efficiently without relying heavily on manually written rules or the extraction of specific features. NLP, a subfield of AI, helps extract meaningful insights from unstructured textual data such as project reports, schedules, and contracts, which are often rich sources of information in construction projects [16]. Through pretraining and fine-tuning, large language models can demonstrate excellent performance [17], significantly enhancing the efficiency and accuracy of information processing, particularly in areas such as unstructured text processing, information extraction, and intelligent decision support, as well as the ability of the construction industry to use the language in a variety of ways. Information extraction and intelligent decision support provide the industry with unprecedented technological means [18].

3.1.2. Deep Learning and Machine Learning Technologies

Deep learning and machine learning are core technologies driving the development of intelligent construction. Machine learning is a subfield of artificial intelligence that enables computers to solve problems by observing a given dataset and generating models based on input data. This differs from traditional programming in that machine learning does not rely on explicitly written rules but instead learns from data to generate predictive models for making predictions on unseen data [18]. Common machine learning methods include supervised learning, unsupervised learning, and reinforcement learning, which are used for tasks such as classification, clustering, and optimizing strategies through interaction with the environment. Deep learning, a subfield of machine learning, is characterized by the use of multiple layers of simple, adjustable computational elements, performing complex tasks by stacking neural network layers [19]. Deep learning methods can typically handle more complex and larger-scale data and outperform traditional shallow learning models in many applications [20].

In smart construction, commonly used deep learning and machine learning techniques include convolutional neural networks (CNNs), recurrent neural networks (RNNs), generative adversarial networks (GANs), and reinforcement learning. CNNs are primarily used for image processing, enabling automatic identification of safety hazards at construction sites [21]. For example, OP-Net effectively identifies various targets in complex construction site images, providing strong support for safety monitoring and progress management at construction sites [22]. When dealing with time-dependent information, RNNs perform better than other forms of deep learning [23], making them suitable for real-time prediction tasks in construction project quality management [23]. GANs excel in adversarial generation, especially in synthesizing images, text, and table data, and can generate multiple design alternatives in design optimization [24,25]. Reinforcement learning is used in controlling construction robots and automated construction equipment, such as assigning different roles to robots (e.g., bricklaying, material handling, and quality inspection) to improve task efficiency and continuously optimize operational strategies through interaction with the environment [26]. Additionally, transfer learning and GNN have also been promoted in cross-disciplinary applications in construction projects, with the former accelerating model training through knowledge transfer and the latter showing advantages in building structure analysis and construction resource optimization [27,28].

3.2. Application Scenarios of LLMs in AECO Industry

The AECO (architecture, engineering, construction, and operation) industry encompasses multiple interrelated phases, each with its unique tasks and challenges. Large language models (LLMs), such as GPT-based systems, offer transformative capabilities across this lifecycle—from early-stage design ideation to long-term facility management. The following Table 1 outlines the key roles LLMs can play at each phase. In architecture, they assist with conceptual exploration, sustainability evaluation, and communication with stakeholders. In engineering, LLMs support technical documentation, compliance checking, and design automation across disciplines like structural or electrical systems. Within the construction phase, LLMs enhance safety monitoring, automate reporting, assist scheduling, and even contribute to robotics integration. Finally, in the operation phase, these models facilitate intelligent maintenance, energy optimization, and predictive analysis. By aligning LLM capabilities with each AECO stage, this framework illustrates how generative AI can support a more integrated, efficient, and knowledge-driven built environment lifecycle.

3.2.1. AI-Driven Information and Interaction Systems

(1). Building Information Modeling (BIM)

Traditional BIM information retrieval is characterized by the use of fixed query formats, such as structured query languages or predefined search parameters. Users are required to have a comprehensive understanding of the BIM data schema and the specific syntax needed to extract information [80]. This method often leads to inefficiencies, as it lacks flexibility and is not intuitive for users who may not be technically proficient. The integration of building information modeling (BIM) and AI-driven question-answering systems is enhancing the intelligence level of information retrieval in construction projects. Lin [79] proposed an intelligent question-answering system based on the BERT (Bidirectional Encoder Representations from Transformers) model, which can understand and parse complex data in BIM to achieve efficient querying and interaction with construction project information. The system uses NLP technology to recognize user query intentions, providing precise BIM data support and enhancing the convenience and accuracy of information retrieval. Zheng [29] introduced a virtual assistant framework based on dynamic prompting technology, combining Transformer language models to dynamically adjust query prompts, improving the accuracy and relevance of BIM information searches. The core of this approach is the use of the contextual understanding capabilities of LLMs, which continuously optimize query content based on user input, making information retrieval more targeted and intelligent. In BIM systems, real-time interaction is crucial for enhancing design efficiency. Fernande [30] developed a virtual assistant named DAVE (Digital Assistant for Virtual Engineering), which uses the GPT model to provide an intelligent interface for BIM systems. This assistant supports both text and voice interaction, enabling real-time querying, updating, and adjustment of BIM data, offering architects and engineers a more intuitive and efficient user experience.

The complexity of building regulations and standards makes BIM compliance checks a challenging task. Chen [31] proposed an automated BIM compliance-checking framework based on large language models and deep learning. This method combines NLP and ontology technology to automatically parse and validate regulatory requirements in BIM models, improving the efficiency and accuracy of compliance checks. Furthermore, the framework supports hierarchical classification of regulations (single-layer, double-layer, and triple-layer information structures), better aligning with various types of compliance needs.

(2). Expert Systems and Human–Computer Interaction

With the widespread application of large language models in the architecture, engineering, and construction industry, their role in expert systems and human–computer interaction has been increasingly prominent. LLMs, when combined with VR, BIM, robot operating systems (ROSs), and human-in-the-loop learning, provide the construction industry with more intelligent and automated solutions. This section will focus on the applications of LLMs in multimodal interaction, education and training, and regulatory compliance annotation. Park [62] explored the integration of LLMs with multimodal virtual reality interfaces to support collaborative human–machine construction tasks. The system proposed in this study combines user interaction channels, ROS, BIM, and game engines such as Unity, and generates interactive objects through Rhino and Grasshopper applications to enhance the intuitiveness and effectiveness of human–computer interaction. In task management, the system employs a bidirectional communication mechanism supported by LLMs, capable of detecting and correcting potential errors, thereby improving the accuracy and efficiency of construction tasks. Its chat system uses prompts generated by GPT, which include two contexts and four task instructions, ensuring effective communication between human operators and robots.

(3). Knowledge Sharing and Educational Integration

In the field of architectural education and knowledge sharing, Maalek [81] studied the integration of GAI with problem-based learning (PBL) to enhance the effectiveness of digital architecture courses. The study found that interaction with LLMs not only assists with code generation, lowering the programming threshold, but also subtly cultivates project management skills and improves students’ abilities to solve open-ended problems. Additionally, this approach helps students view project development as an opportunity for learning and growth, stimulating their innovative capabilities. This suggests that the application of LLMs in architectural education has vast potential, especially in code generation, problem solving, and knowledge sharing. Al-Turki [35] proposed an LLM-based human-in-the-loop learning system to achieve efficient RASE (requirements, applicability, selection, and exceptions) tagging. The study explored methods such as few-shot learning, fine-tuning learning, and progressive active learning to automate the parsing and annotation of regulatory texts, effectively simplifying the extraction of regulatory rules and improving the efficiency of compliance audits. This research demonstrated the potential of LLMs in architectural regulatory compliance and indicated that the human–computer collaboration model could enhance the intelligence of compliance tasks.

In large-scale urban development projects, chatbots can engage in design dialogues with stakeholders, breaking traditional communication barriers by automatically collecting and organizing conversation data and converting it into a design requirements list. In the study by Dortheimer [32], a chatbot system was created to facilitate design dialogues in large-scale urban development projects. The system can hold design conversations related to specific architectural projects and convert those discussions into a requirements list, automating and streamlining the information exchange between stakeholders and designers. Al-Turki propose utilizing this technology to enhance design discussions in large-scale urban development projects by developing chatbot systems that automate and streamline information exchange between stakeholders and designers [35]. Compared with traditional surveys, chatbots provide a more engaging and interactive experience, making participants enjoy the communication process more, thereby increasing user satisfaction [82]. This encourages greater participation in the design process, providing more ideas and suggestions for the project and fostering public involvement in design projects.

(4). Data and Corpus Development

The application of large language models in the architecture, engineering, and construction AEC field also relies on high-quality corpus support. Zhong and Goodfellow [36] studied the feasibility of pretraining domain-specific language models on construction management system corpora. Their research indicated that the application of Transformer models in the construction management field is still in its early stages. However, as task complexity increases, these models can more effectively capture contextual information, thereby improving the accuracy of information extraction and the precision of task execution. Similarly, Zheng [37] proposed a pretrained language model for information retrieval tasks in the AEC domain. This work systematically explored the impact of domain-specific corpora and various transfer learning techniques on the performance of deep learning models in information retrieval tasks and proposed a pretrained domain-specific language model for the AEC field. Gao [38] introduced an intelligent knowledge management solution for full-process engineering consulting, which includes data collection and preprocessing, engineering consulting knowledge model description, and knowledge joint extraction. The LF-CASREL model integrates deep learning and natural language processing technologies, enabling simultaneous extraction of entities and relationships from full-process engineering consulting knowledge texts.

3.2.2. Engineering Management and Construction Optimization

(5). Contract Management

Contract review and risk management are critical aspects of the construction preparation phase, involving the analysis of construction specifications, contract clauses, and potential legal risks. In recent years, the application of large language models and NLP in this field has made significant progress. Moon [71] proposed a BERT-based contract risk clause classification method, utilizing NLP techniques to identify seven categories of contract risks, including payment, time, procedures, safety, roles and responsibilities, definitions, and references. The study trained a model on 2807 clauses extracted from 56 construction specifications, achieving a validation accuracy of 0.889 and an F1 score of 0.934, significantly enhancing the automation of contract review. This study demonstrates that BERT can be used to automatically screen potential contract risks in construction specifications, improving decision-making for project stakeholders. Wong [72] further introduced a knowledge-augmented LLM for construction contract risk identification (CCRI). This method integrates a contract knowledge base with a two-stage prompting approach, enabling expert-level contract risk assessment without fine-tuning (tuning-free). The study found that the model can perform reasoning based on an external knowledge base, effectively improving the accuracy and interpretability of contract risk detection. Compared with traditional methods, this approach makes the contract review process more intelligent while reducing the time costs associated with manual review. Kazemi [72] focused on the automatic detection of construction contract risk clauses written in complex script systems, such as Arabic and Persian. Their research applied NLP models to extract fragmented engineering consulting knowledge and integrate it into contract management. This method plays a key role in knowledge management for whole-process engineering consulting, offering new perspectives for multilingual contract management in international construction projects.

(6). Construction Scheduling and Task Optimization

Construction project management encompasses various aspects, including scheduling, resource allocation, and cost control. The application of large language models in construction management, particularly in construction scheduling and task optimization, is demonstrating significant potential. You [63] explored the use of ChatGPT in construction project scheduling and proposed the RoboGPT framework for automating assembly sequence planning (ASP). Their study indicated that ChatGPT can be leveraged for reasoning and optimizing construction scheduling, enhancing the automation of task planning while reducing the need for manual intervention. He [44] investigated constraint management in construction planning using GPT-4 and introduced a constraint analysis method based on BERT and GPT-4. This approach analyzes construction meeting records to automatically identify different stakeholders’ concerns—for instance, management focuses on global constraints, while foremen and subcontractors are more concerned with specific operations. Additionally, the method can extract the root causes of construction constraints and provide AI-driven optimization suggestions, thereby improving communication and decision-making efficiency in construction management.

Supply chain management in construction projects involves numerous suppliers, contractors, and logistics operations, making risk control crucial for project progress and cost management. The application of LLMs and Transformer-based technologies in construction supply chain risk management (CSCRM) has significantly enhanced the efficiency of risk identification and information extraction. Shishehgarkhaneh [48] investigated the use of Transformer-based named entity recognition (NER) techniques for supply chain risk management in the Australian construction sector. Their method enables the automatic extraction of supply chain risk-related entities from large-scale news data, providing decision-making support for construction firms and regulatory bodies. The study demonstrated that integrating advanced NLP models, such as BERT, into supply chain risk management systems can significantly improve the security and risk control capabilities of the construction industry’s supply chain.

(7). Construction Safety Management

Construction safety is a critical component of project management, involving the identification, analysis, and prevention of risks on construction sites. With advancements in artificial intelligence, researchers have leveraged natural language processing, deep learning, and large language models to enhance safety information management, accident retrieval, and prediction capabilities, ultimately reducing construction risks and ensuring worker safety. Shin [55] explored the optimization of construction accident information classification and similar accident retrieval using the KLUE-BERT model. Their study fine-tuned KLUE-BERT for better adaptation to the Korean language and applied it to: (1) Information Classification: using named entity recognition (NER), the model categorized accident report details into 18 specific types (e.g., accident time, weather conditions, construction type). (2) Accident Retrieval: employing semantic textual similarity (STS) tasks, the system retrieved the most similar historical accident cases to assist in construction safety management. Tian [56] proposed an intelligent question-answering method for construction safety hazard identification based on deep semantic mining. Their approach integrated BERT, bidirectional gated recurrent units (BiGRUs), and self-attention mechanisms into a deep learning network to extract semantic features from construction safety texts. A Siamese network was then used to match safety issues with management measures, enhancing the accuracy and automation of hazard identification. Liu [57] examined how scene graphs and information extraction (IE) techniques could be combined to automatically identify construction hazards. They introduced a BERT-based information extraction model capable of processing regulatory texts and extracting key information to detect safety violations, improving regulatory oversight on construction sites. Smetana [58] investigated the application of LLMs in highway construction safety, demonstrating their effectiveness in classifying construction accidents and improving adaptability to different construction scenarios. By automatically analyzing common safety hazards in highway construction—such as high-speed traffic, heavy machinery operations, and adverse weather conditions—LLMs significantly enhanced safety measures. Zhang [49] proposed an automatic hazard identification method that integrates construction scene graphs with BERT, referred to as the C-BERT network. This method first uses computer vision to extract scene information from construction sites, including entities, attributes, and interactions. It then incorporates construction safety regulations for hazard inference. Zhong [54] explored a construction safety video analysis approach based on visual attention mechanisms. Their method automatically calculates the importance distribution of information in construction videos, extracting key objects, relationships, and attributes to enhance semantic recognition in safety monitoring systems. This approach improves hazard detection on construction sites and provides real-time safety alerts. Yoo [60] leveraged GPT for construction accident prediction and proposed a saliency visualization method. By fine-tuning a GPT model, they classified unstructured text from construction accident reports into various categories (e.g., crushing, falls, collisions). A saliency analysis was then conducted to quantify the influence of individual words on the prediction outcome, identifying key factors contributing to accidents and improving AI model interpretability. By analyzing construction reports and site data, this approach enables proactive risk identification, shifting safety management from reactive responses to preventive measures.

(8). Risk Management

Construction sites generate vast amounts of data, including progress records, safety inspection reports, and material usage information. Efficient collection, analysis, and generation of construction reports have become key priorities in modern construction information management. Pu [64] proposed the AutoRepo framework, which utilizes drones and multimodal LLMs to achieve automated data collection and report generation for construction sites. The system’s key features include: (1) automated drone inspections to capture site images and sensor data; (2) multimodal LLMs for automatic construction report generation, ensuring data analysis accuracy; and (3) an expert review mechanism to enhance the reliability and credibility of reports. Li [39] investigated an intelligent compliance verification method for construction plans by integrating knowledge graphs and LLMs. This method relies on information retrieval, semantic understanding, and computational reasoning to automatically verify whether construction plans comply with building regulations and industry standards while providing real-time compliance recommendations, achieving an accuracy of up to 72%. Their study demonstrated that large language models could effectively facilitate construction compliance management, reducing risks caused by non-compliant construction plans.

As construction projects become increasingly complex and uncertain, risk management is progressively shifting towards data-driven and intelligent approaches. In recent years, artificial intelligence technologies, particularly LLMs and deep learning, have provided new solutions for identifying, assessing, and mitigating construction risks. Lee [50] proposed a construction risk assessment knowledge base using BERT and graph models, which leverages natural language processing techniques to extract predefined knowledge from unstructured construction data and construct an entity–relationship-based risk assessment knowledge base. This method effectively manages and extracts risk-related information to support construction safety management. Their study highlighted that this knowledge base could be used to identify risk factors such as machinery, equipment, materials, and construction methods, thereby enhancing the intelligence level of construction safety management. Zhou [51] explored the integration of knowledge and deep learning to intelligently generate risk mitigation measures for subway construction. They employed the K-BERT model, embedding domain-specific knowledge from subway construction into the model in the form of triplets to enable the intelligent generation of construction risk response measures. Their study improved the efficiency of risk mitigation and introduced new perspectives for intelligent risk management. Nyqvist [53] examined the capabilities of GPT-4 in construction project risk management (CPRM), with a focus on its performance in risk identification, analysis, and control. Cao [52] proposed a risk evolution analysis method based on textual semantics. By extracting risk factors and their relationships from construction accident case texts, they constructed a risk coupling evolution knowledge graph and applied text matching and path reasoning algorithms to enable intelligent inference and quantitative evaluation of construction risks. Their study contributed to identifying potential risk patterns in construction projects, improving risk prediction and management capabilities.

3.2.3. Operations and Sustainability

(9). Energy Management

In building energy management, LLMs have been used to optimize building energy models, reduce modeling workload, and ensure high accuracy: “The platform was tested with a number of modeling cases, demonstrating reduced modeling efforts by over 95% while ensuring 100% accuracy in building energy models’ creation” [68]. This indicates that LLMs not only enhance the efficiency of knowledge sharing but also contribute to the optimization of energy management and promote sustainability. In the field of building energy performance analysis, Forth and Borrmann [33] explored how LLMs and semantic text similarity techniques can be used to optimize BIM modeling. The study showed that semantic matching can more accurately assign thermal performance parameters to spatial elements in BIM models, thus improving the accuracy of building energy performance simulation (BEPS). Additionally, the study demonstrated how this method can be integrated into decision support tools to assist project teams in more effectively optimizing building energy efficiency.

(10). Material Innovation

Tian [43] proposed a large language model-based method for steel structure design and developed a language model called SteelBERT. SteelBERT, pretrained on literature data related to steel materials, can capture material knowledge and achieve high-precision performance predictions. For example, the model can predict the mechanical properties of 18 new types of steel using text sequences and the context of the manufacturing process, with prediction accuracies of 78.17% (YS), 82.56% (UTS), and 81.44% (EL), significantly outperforming traditional methods. Additionally, the study also optimized the manufacturing process and achieved the development of new materials that surpass the performance of the reported 15Cr austenitic stainless steel.

(11). Geotechnical Engineering

Kim [41] proposed a ChatGPT-MATLAB framework for numerical modeling in geotechnical engineering applications. This framework communicates geotechnical engineering problems, such as two-dimensional seepage analysis, slope stability analysis, and X-ray CT image phase segmentation, to ChatGPT via descriptive prompts, which then automatically generates MATLAB code. This approach reduces the manual coding workload traditionally required and achieves precise numerical modeling automation by clearly defining geometric parameters, initial conditions, boundary conditions, and calculation methods. Kumar [42] explored the potential of GPT in geotechnical engineering, suggesting that ChatGPT, as a knowledge discovery and reasoning tool, can accelerate the application of data-driven methods and computational modeling. For example, the research demonstrated that GPT could classify fine-grained soils based on their liquid and plastic limits and evaluate the safety of gravity retaining walls.

(12). Operations and Maintenance

During the building operation and maintenance (O&M) phase, efficient information retrieval, rapid problem identification, and intelligent execution of maintenance tasks are crucial for improving the efficiency and accuracy of building management. The application of large language models in this phase mainly focuses on information retrieval and enhancement, real-time issue identification during building meetings, building defect analysis, data point labeling in building management systems, and augmented reality (AR)-assisted maintenance tasks. Lee [61] compared the performance of retrieval-augmented generation (RAG) and fine-tuned LLMs in building safety management knowledge retrieval. They found that the RAG model, combining GPT-4 and a building safety knowledge graph, more effectively extracted relevant knowledge and generated preventive measures when summarizing accident reports. In building meeting management, Chen [45] proposed an LLM-supported framework (Meet2Mitigate, M2M), which uses speech recognition technology for speaker labeling and combines LLMs for abstract summarization and problem extraction. This method can effectively capture key issues discussed during building meetings and use RAG technology to access building knowledge databases, offering best practice suggestions for the issues raised. Furthermore, LLMs can generate detailed action items, improving the executability of meeting decisions.

Additionally, defect analysis in residential buildings is essential for improving construction quality. Shooshtarian [67] explored the application of NLP in residential building defect analysis, showing that LLMs like BERT and GPT-3 perform exceptionally well in named entity recognition tasks and can be used for automatic classification and analysis of building defect data. Moreover, LLMs’ application in building maintenance tasks can be further optimized through AR technology. Xu [69] studied the application of ChatGPT in AR-assisted maintenance tasks, proposing an LLM-based text-to-action conversion system. Based on user interaction with the AR interface, the system utilizes LLM’s language understanding and reasoning capabilities to determine appropriate steps and present them to the user as virtual symbols superimposed on the relevant environment and facilities, offering real-time, context-aware guidance and optimizing O&M task processes.

3.3. Data and Information Processing

3.3.1. Data Collection and Preprocessing

In the field of construction engineering, large-scale construction projects generate vast amounts of data from diverse sources and in various formats. These include, but are not limited to, CAD drawings (vector data), construction logs (sequential text), BIM models (three-dimensional spatial data), IoT sensor data (time-series signal streams), and drone aerial imagery (unstructured visual data) [29,30,79]. These data sources differ not only in format but also in quality, often exhibiting the following issues: (1) data redundancy and fragmentation, such as duplicated procedural entries in construction logs; (2) data quality deficiencies, including sensor noise, manual entry errors, and format inconsistencies; and (3) semantic barriers, where variations in data encoding rules and measurement standards exist across different systems [33,83]. To improve data quality and ensure the accuracy and consistency of subsequent information processing, data preprocessing is a critical step involving data cleansing, standardization, and integration.

Data cleansing serves as the foundation of data preprocessing, aiming to eliminate noise, fill in missing values, and correct or remove incomplete or inaccurate information, thereby enhancing data reliability and integrity. Data standardization ensures that information from various sources can be uniformly managed and analyzed. Given that engineering data are often collected from multiple platforms and systems with varying measurement standards across different entities, preprocessing—such as cleansing and standardization—is essential for maintaining data quality and consistency [66]. Standardization ensures compatibility and interoperability under a unified framework, laying the groundwork for subsequent information extraction and analysis. Lastly, data integration is a key step in achieving efficient data management, allowing structured, semi-structured, and unstructured data to be consolidated into a unified data model, thereby enhancing usability and system compatibility. For instance, one study integrated multi-source data—such as meeting transcripts and project management records—to enable cross-validation of information across different dimensions, ultimately improving the accuracy of data-driven decision-making [44].

3.3.2. Information Extraction and Knowledge Graph Construction

After completing data preprocessing, information extraction becomes a critical step in the data processing pipeline. Leveraging deep learning-driven multimodal information extraction frameworks, it breaks through the limitations of traditional structured data analysis, enabling precise capture of engineering semantics. It allows for the extraction of specified types of entities, relationships, events, and other factual information from vast amounts of construction engineering data, converting them into structured data outputs [84]. This process relies on NLP, computer vision, and other machine learning methods to integrate multiple types of data, including text, images, audio, and video. Multimodal information extraction maximizes the semantic associations between different modalities, enhancing the effectiveness of information extraction [84].

The first step involves multimodal entity recognition, where named entities, their types, and corresponding boundary box locations are extracted from image–text pairs, enabling the automatic identification of key entities with real engineering significance. The next step is relationship extraction, which uses attention mechanisms and graph convolution networks to capture the complex relationships between entities [85,86,87,88]. This method uncovers deeper connections between entities, further exploring logical links between them, thus transforming isolated data points into an interconnected knowledge network. Relationship extraction can identify causal relationships, temporal relationships, spatial relationships, and more, providing support for intelligent modeling and knowledge reasoning [89].

Lastly, attribute extraction focuses on extracting detailed parameters that describe the characteristics of various entities. It involves identifying predefined categories of entities, relationships, and event instances from natural language text and extracting the relevant attributes (parameters) of these identified entities, relationships, or events [90]. For example, attribute extraction can identify climatic data, building operational data, and the physical parameters of a building within a building management system, covering various aspects of building operations—from climate conditions and energy consumption data to the physical characteristics of the building itself. This process provides high-quality foundational data for engineering optimization [70].

Constructing a knowledge graph involves integrating the recognized entities, attributes, and their interrelationships [74]. The knowledge graph can comprehensively depict the full lifecycle information of a construction project, such as component parameters during the design phase, timelines during the construction process, material usage, and risk factors [91]. This process not only improves the visual representation of data but also provides strong support for intelligent question answering, automatic reasoning, and predictive analysis. Additionally, structural representation makes complex information in construction project management more logical and hierarchical [92]. Framework extraction techniques organize and model engineering data to form clear logical structures [93], such as temporal relationships in project progress, causal chains in risk management, and optimization plans for resource scheduling. This structured way of expressing information enhances the operability of data, enabling engineers and managers to quickly understand and utilize it effectively.

3.4. Core Components and Key Technologies of the Toolchain

With the widespread application of large language models across various industries, it has become crucial to build efficient and flexible toolchains and achieve seamless integration between systems. A toolchain refers to a set of integrated tools and technologies that work together to complete specific tasks or workflows. In the context of LLM applications, a toolchain includes components such as data processing, model training, model inference, APIs, plugins, and more. As shown in Figure 4, these components collaborate within a particular business scenario to improve efficiency, accuracy, and intelligence.

System integration is the process of seamlessly connecting different technologies, software, and hardware components. It involves embedding LLM technology into existing software platforms, applications, and workflows, such as BIM, CAD, etc., to enable cross-system functional collaboration. For instance, Kang [34] proposed an earthwork network architecture that utilizes LLMs to process CAD drawing data and convert it into structured, trainable datasets. This method overcomes the challenge of handling unstructured data in traditional earthwork volume estimation, improving prediction accuracy and computational efficiency. It allows for more precise predictions of earthwork areas and automatically extracts features for earthwork volume calculations, significantly enhancing construction project planning capabilities.

This integration of LLMs with existing systems like BIM and CAD not only improves the efficiency of specific tasks but also enables more advanced, data-driven decision-making in construction management and design, thus supporting smarter and more automated workflows.

LLMs serve as foundational technology, and through toolchains, they integrate the capabilities of LLMs with the functions of traditional software or platforms, enhancing the system’s intelligence, especially in data processing, analysis, and automation tasks. APIs (application programming interfaces), plugins, and customization are key means by which LLMs are integrated into existing software ecosystems. These tools allow LLMs to seamlessly combine with various software tools, highly complex workflows, and industry-specific systems, providing users with flexible, efficient, and intelligent solutions. This process involves not only technical challenges but also ensures compatibility and collaboration between different platforms and tools. For example, Pu [64] proposed the AutoRepo framework, which uses drones and multimodal LLMs for automatic data collection and report generation at construction sites. The system features: (1) drone-based automatic inspections, collecting images and sensor data from the construction site; (2) a multimodal LLM for automatically generating construction reports to ensure data analysis accuracy; and (3) an expert review mechanism to enhance report reliability and credibility.

An API is the main way through which LLMs interact with external systems, allowing different software platforms to call upon the capabilities of LLMs for tasks like text processing, data analysis, and automation. Raiaan [15] mentions that the OpenAI API supports tasks such as programming, Q&A, and text analysis, and enhances automated text processing via the Chat Completions API, offering interactive dialogue and function calls. The Hugging Face API provides access to over 150,000 machine learning models, widely used for text analysis, speech recognition, and image processing, while supporting rapid model switching. Google Cloud API and Microsoft Azure Language APIs offer cloud-based NLP solutions, including sentiment analysis, text auditing, entity recognition, and text summarization, facilitating easy integration into existing systems. Plugins are another important form of LLM integration into existing software tools. Compared with APIs, plugins allow users to introduce LLM capabilities without altering the main software architecture, thereby enhancing the system’s intelligence.

Although LLMs possess strong general language processing capabilities, industry applications often require customized adjustments to meet specific business needs. Enterprises can optimize LLMs through fine-tuning and prompt engineering, making them more suitable for the required scenarios. In the construction industry, LLMs can be specially trained to understand architectural terminology, construction standards, and industry regulations, allowing them to more accurately interpret project documents and generate compliance reports. Some companies are exploring the construction of private large models based on LLMs, which can not only integrate the company’s internal knowledge base but also enhance “domain-specific queries” and “private knowledge bases” through RAG, collecting and labeling industry-related datasets to better understand and generate domain-specific terminology and standards.

4. Existing Challenges

4.1. Data Layer

4.1.1. Data Quality and Corpus Construction

While the construction of domain-specific databases is a focus in many fields, the construction industry faces more complex data types compared with sectors like healthcare and law. For instance, data types in construction include BIM data, construction logs, contract texts, regulatory documents, and sensor data. These data come in diverse forms such as text, images, 3D models, and videos, making it difficult for traditional NLP techniques to handle them uniformly. Existing large language models are primarily trained on general text data (e.g., Wikipedia, Common Crawl), but high-quality, structured corpora specific to the construction industry are scarce. The development of domain-specific corpora and the open sourcing of pretrained models are crucial to advancing NLP applications in the AEC industry, but there is currently a lack of publicly available resources [37]. Additionally, there is no unified semantic labeling or public training dataset in the AEC field, resulting in a heavy reliance on manual data annotation, which is both costly and time-consuming [37,51]. This presents a significant bottleneck in applying LLMs to the AEC industry. Zhong [36] proposed a domain-specific language model pretrained on a construction management system corpus, mainly extracted from academic papers. Their research also mentioned specific datasets for construction management, such as inspection report datasets for infrastructure condition prediction and building code datasets for automatic compliance checks. However, the scope of this approach is still limited by the specific data sources, and cross-project data standardization issues have not been effectively addressed [35]. Moreover, much of the data in the AEC industry is not standardized, structured text but rather unstructured information such as construction meeting minutes, field reports, and risk assessment reports. Lee [40] found that LLMs had a low accuracy in extracting relevant knowledge from unstructured text during the construction risk assessment knowledge base construction, directly affecting the model’s subsequent automated risk assessment and decision support capabilities. The application of current LLMs in the AECO industry still has clear limitations, and improving the processing of unstructured data, optimizing corpus construction processes, and achieving efficient cross-project data sharing remain important challenges for future research.

4.1.2. Data Heterogeneity and Interoperability Issues

The AECO industry is characterized by a high degree of data heterogeneity, stemming from the diverse range of tools and formats used across project phases. Major software platforms such as Revit, ArchiCAD, and Bentley Systems employ proprietary data schemas and closed ecosystems, which introduce substantial challenges in terms of data conversion, interoperability, and information fidelity [75]. These interoperability barriers often result in semantic misalignment, data loss, or conversion errors during model exchange, significantly affecting collaborative workflows.

Saka [75] emphasized that large-scale projects typically engage multiple structural consultants and domain experts who utilize different analytical and modeling applications. The absence of seamless interoperability between these tools hampers effective data sharing, consistency in interpretation, and timely decision-making. These problems are exacerbated when integrating LLMs, which require structured, semantically consistent data inputs to support tasks such as reasoning, compliance checking, and automated decision support.

Although the Industry Foundation Classes (IFC) standard provides a neutral format for data exchange [94], it does not fully resolve semantic discrepancies or address the deep contextual understanding required by LLMs. For instance, compliance verification tasks in BIM may span single-layer data, double-layer relationships, and triple-layer system rules. Current LLMs often lack the ability to construct coherent logical mappings across these layers, limiting their effectiveness in high-stakes applications like code compliance, structural optimization, and lifecycle management [31,95].

Moreover, while preliminary attempts to integrate LLMs with domain-specific platforms—such as the use of ChatGPT alongside MATLAB or Dynamo—demonstrate conceptual feasibility [41], practical implementation faces nontrivial barriers. These include issues related to API availability, real-time bidirectional data flows, and the lack of standard protocols for model-level interaction between generative AI and traditional AEC software environments like Revit or Navisworks. As such, LLMs currently operate largely in isolation or in postprocessing modes, rather than as real-time, embedded agents within existing AECO digital ecosystems.

In summary, overcoming interoperability challenges is crucial for unlocking the full potential of LLMs in AECO applications. Future research must prioritize the development of standardized, semantically enriched interfaces and middleware solutions that bridge the gap between probabilistic LLM reasoning and deterministic domain software logic.

4.2. Technical Layer

4.2.1. Collaborative Limitations in Technology Integration

The construction industry already faces heterogeneous data sources, which include structured knowledge (such as ISO 19650 standards [96] and building regulations) as well as unstructured texts (like construction logs, contract terms, and meeting records). One of the major challenges in current research is how to effectively integrate LLMs with knowledge graphs (KGs) for efficient reasoning and dynamic knowledge updates. Knowledge graphs, as a core technology for structured knowledge representation, have significant potential for applications in the AECO sector, such as risk management, regulatory compliance checks, and building operation and maintenance. However, knowledge graphs emphasize explicit logical reasoning, while LLMs rely on implicit probabilistic distributions for inference, leading to fundamental compatibility issues between the two approaches.

Lee [50] explored a hybrid BERT+KG architecture for construction risk assessment and found it effective for basic risk identification. However, the system exhibited limitations in handling semantically complex, multi-factor risks—such as fall hazards, which depend on factors like safety measures, weather, and personnel training—indicating insufficient reasoning capability under compounded conditions.

In a broader context, Park Somin noted the technical and systemic challenges in integrating LLMs with diverse platforms such as the robotics operating system (ROS), building information modeling (BIM), and game engines. Their findings underscore the current fragmentation and the need for unified frameworks to facilitate seamless interoperation among these technologies.

4.2.2. Real-Time and Accuracy Issues in AI Interaction

In VR human–machine collaboration and smart construction scenarios, efficient interaction with AI is crucial for enhancing construction automation. However, current LLMs still face significant challenges in terms of real-time performance and accuracy when interpreting construction-related information, generating task instructions, and planning robotic actions. Complex models often require substantial computational resources, leading to higher latency. This cumulative latency effect is particularly evident in large construction projects, potentially causing delays in construction processes and even impacting safety.

In addition to computational delays, AI also faces issues of interaction efficiency in BIM-driven robotic control tasks. Park et al. [62] proposed a conversational VR interface integrating multimodal interaction, including voice and controller inputs, to enhance intuitive communication between construction workers and robots. Their system demonstrated low workload and high usability with succinct command inputs.

Luo [65] introduced SafePlan, a multi-component framework that combines formal logic and chain-of-thought (CoT) reasoning to enhance the safety of LLM-based robotic systems. Their results showed that SafePlan outperforms baseline models by leading to a 90.5% reduction in harmful task prompt acceptance while still maintaining reasonable acceptance of safe tasks.

4.3. Application Layer

4.3.1. Task Specificity: Weak Generalization Ability in Complex Scenarios

Although LLMs have demonstrated strong natural language understanding and generation capabilities in many industries, such as healthcare and law, their application in the AECO industry still faces numerous challenges. In practical applications, the generalization ability of models is crucial [48]. The construction industry is highly complex and involves multiple stages, including structural design, construction management, regulatory compliance, and intelligent operation and maintenance. However, existing LLMs are not yet mature in these tasks and struggle to fully address the complexity of the industry. For example, hazard identification on construction sites requires consideration of the interactions between different construction components, but existing computer vision methods fail to generate interaction-level scene descriptions [49]. Moreover, construction safety planning involves highly contextual tasks, as different construction environments (such as high-altitude welding, underground excavation, and tunnel digging) pose different safety risks. Existing LLMs are difficult to precisely adapt to specific construction scenarios, leading to safety guidance information that is too generalized and difficult to directly apply to specific working conditions [59]. Although RAG (retrieval-augmented generation) can alleviate hallucination issues by enriching user queries with relevant information from external knowledge sources, this method has limitations, especially when specific domain or scenario knowledge is required. It may not ensure safety and compliance [97].

4.3.2. Multi-Role Collaboration: Conflicting Interests Leading to Information Bias

Typically, the goal of different parties involved in a construction project is to maximize their own interests, often overlooking the goal of achieving a relatively balanced optimization of the needs of all stakeholders and the overall best performance of the project. Tae Keun Oh [98] pointed out that different stakeholders (such as clients, designers, contractors, and safety managers) have varying concerns and expectations regarding safety measures, and these differences in goal preferences make it difficult for LLMs to generate construction safety advice that satisfies the needs of all stakeholders. LLMs often rely on existing data and general standards when generating construction safety suggestions, lacking a deep understanding of the specific project environment, the needs of all parties, and the actual trade-offs involved. These vast datasets may contain human biases, and therefore, decision models trained on such data may inherit these biases, potentially leading to unfair and incorrect decisions [99]. As a result, the safety guidance generated by LLMs may tend to favor one party’s perspective, and this inconsistency in goal preferences makes it challenging for the generated suggestions to meet the needs of all stakeholders simultaneously, which in turn affects the collaborative nature and effectiveness of construction safety management. Furthermore, research by He [44] reveals significant differences in the constraints discussed by different professional roles in construction planning meetings, where management team members tend to balance various constraints, while foremen and subcontractors focus more on practical issues. In large-scale construction projects, traditional clients are usually government agencies or developers rather than end-users, making meaningful dialogue with thousands of potential stakeholders extremely difficult [32].

4.3.3. Economic Constraints: High Computational Resources and Maintenance Costs

The application of LLMs in the construction industry faces significant computational resource consumption. Training and running a generative AI model add extra costs to projects, posing challenges to its adoption [76]. Compared with traditional NLP tasks, AECO applications often involve complex geometric modeling, real-time interactions, and task planning, placing stricter demands on computational resources. Wu Bing yang [100] pointed out that LLMs require significantly higher GPU memory when handling complex tasks than traditional NLP models. Training LLMs requires substantial computational power and expensive hardware (such as GPUs or TPUs), limiting their widespread application in the building energy sector [68].

Furthermore, integrating BIM with equipment engines further exacerbates computational burdens. For example, in VR environments, real-time rendering of building information models, combined with human–computer interaction and task guidance, requires simultaneous processing of natural language understanding, graphical computation, and physical simulation—placing extremely high demands on hardware [62]. This not only increases the initial deployment costs but also raises IT infrastructure maintenance requirements for enterprises.

In addition, to adapt to the dynamic changes in construction regulations, real-time monitoring of construction environments, and adjustments to industry standards, LLM applications in AECO often require frequent fine-tuning and continuous updates. Beyond basic computational resource requirements, the cost of utilizing LLMs is another major hurdle in their deployment within the construction industry. Expenses related to network access, application development, and annual maintenance and updates are prohibitively high, posing challenges for small- and medium-sized enterprises (SMEs), which make up approximately 80% of the industry [75]. The high computational demands and operational costs not only limit the broad adoption of LLMs but also force AECO companies to carefully evaluate the economic feasibility of implementing such technologies.

4.3.4. Human–Machine Trust Gap: Trust and Ethical Risks

In the AECO industry, AI tools such as LLMs are gradually being integrated into various stages of project design, construction management, cost control, and quality monitoring. “Traditional inspection methods are typically expert-based, heavily relying on personal experience and accumulated knowledge [39]”. However, given the high-risk nature of AEC projects, any errors can lead to significant economic and safety consequences, requiring extreme accuracy and reliability in decision-making processes. Despite offering decision support, these AI tools still face a human–machine trust gap and ethical risks, not only in safety management but across multiple critical aspects throughout the project lifecycle.

The “black box” nature of GPT models makes their decision-making process difficult to interpret, impacting trust in the construction industry and necessitating the development of new methods to enhance model explainability [60,77]. While LLMs demonstrate strong capabilities in text generation and information integration, their internal reasoning mechanisms remain opaque to users, and the confidence level of their outputs is difficult to quantify precisely. In the AEC sector—where precision is paramount, particularly in high-risk project environments—decision-makers need AI tools that not only generate reasonable recommendations but also provide insight into the underlying logic and confidence assessments [46]. To effectively integrate AI with traditional engineering expertise, it is essential to improve the explainability of LLMs, ensuring that their outputs align with risk assessment practices and safety standards in engineering, ultimately bridging the trust gap between humans and machines.

Furthermore, the application of large language models must address issues related to data privacy, copyright, intellectual property, and the accountability of model outputs. As the construction industry embraces GPT technology, it also faces significant cybersecurity challenges. Malicious actors could exploit vulnerabilities in interconnected systems, cloud computing, and data exchange, increasing the risks of unauthorized access, data breaches, and cyberattacks [75]. For example, Liu, Mingzhe [68] highlights that integrating LLMs into building energy management systems may pose privacy and cybersecurity risks, such as data leaks and malicious attacks. Given that industry-related technical documentation contains sensitive information, strict data privacy protection is necessary when fine-tuning LLMs [40].

The existing challenges have been categorized into three main dimensions—the data layer, technical layer, and application layer—as illustrated in the Figure 5 below; the subsequent chapter follows this same structural framework to outline corresponding future development directions.

5. Future Development Directions

5.1. Data Layer Innovations

5.1.1. High-Quality Data Construction and Multilingual Adaptation

The effective application of LLMs in the AECO industry critically depends on the development of high-quality corpora and the enhancement of multilingual adaptability. Currently, challenges at the data layer—including inconsistent data quality, lack of standardized domain terminology, and limited multilingual support—significantly constrain the generalization capacity of LLMs across different regions and project scenarios. First, it is imperative to expand and optimize both structured and unstructured AECO-related data resources. These include building information models (BIMs), construction documentation, energy performance assessments, and historical accident reports. Such datasets should undergo standardized annotation and semantic harmonization to enhance the model’s comprehension accuracy and contextual reasoning abilities [29]. Moreover, future research may explore how generative AI models can be leveraged for multilingual translation to support cross-cultural communication and collaboration in global construction projects [78]. This will support the broader applicability of LLMs in globally distributed and multilingual construction environments.

5.1.2. Data Enhancement and Multimodal Integration

Future research should focus on enhancing the comprehensiveness and semantic richness of AECO-related datasets through both content expansion and standardization. For instance, expanding the coverage of BIM and energy performance data will provide models with a more comprehensive material database [33]. Moreover, developing multimodal datasets that integrate images, videos, and text can enhance model understanding of dynamic construction site conditions [54,63]. This direction will not only improve the model’s situational awareness in real-world scenarios but also enhance decision-making quality in complex environments.

Furthermore, Zhang [66] suggest that future research should focus on improving the postprocessing of defect data extraction, including named entity alignment and other techniques, to better understand and optimize the relationship between these factors and model performance. This can help improve information extraction quality and explore efficient data storage and retrieval methods [59].

Future research should prioritize the development and adoption of standardized data schemas and semantic alignment techniques to facilitate seamless data exchange and machine interpretability across systems. Specifically, aligning AECO datasets with open data standards such as IFC, Construction Operations Building Information Exchange (COBie), and domain-specific ontologies can improve interoperability between different platforms and software environments [33,79]. These efforts are crucial for enabling LLMs to reason accurately across tasks such as regulatory compliance checking, design validation, and energy performance prediction.

5.2. Technical Layer Advancements

5.2.1. Technology Integration and System Expansion

In terms of technology integration, previous studies have emphasized the importance of deeply integrating LLMs with existing tools in the AECO industry, such as BIM, CAD, and IoT. For instance, one promising direction is exploring the development of built-in assistants directly within Revit, which would eliminate the need for external data conversion, API calls, and manual database updates, thereby simplifying interactions and improving system responsiveness and accuracy [30]. Additionally, Park [62] highlights the potential of providing models with supplementary contextual data through RAG techniques, enhancing the model’s ability to accurately interpret and classify data. Another noteworthy research direction is the collaborative application of generative AI with AR and VR, which could enable more intuitive and interactive design and construction processes [69]. Moreover, advancing end-to-end workflows that support the full project lifecycle—from design to operation—will facilitate comprehensive lifecycle management in construction projects [29,68]. These integrations and expansions will bring transformative changes to the construction industry, accelerating its transition toward greater intelligence and automation.

5.2.2. Algorithm and Model Optimization

To further explore the application of large language models in the AECO industry, improving overall model performance, reasoning ability, and computational efficiency is crucial for enhancing generalization to new or unseen scenarios. Various optimization strategies have been proposed to achieve this goal. For example, Zheng [29] suggests improving pretraining strategies and using dynamic prompting techniques to enhance the model’s understanding of complex semantics. Additionally, instruction fine-tuning and rule-based validation methods have been recommended to reduce errors in generated content, ensuring accuracy and consistency [31].

For specific tasks such as hazard identification and compliance checks, Luo [65] emphasizes the development of lightweight model architectures to reduce computational costs and improve response speed. Some studies also explore enhancing generalization capabilities by improving embedding techniques and similarity matching methods, which can help models adapt more effectively across various tasks [29,52]. These advancements will play a critical role in increasing the level of automation in the construction industry and accelerating the sector’s digital transformation.

5.3. Application Layer Expansion

5.3.1. User Participation and Scenario Adaptability

As large language models gradually enter practical applications in the construction industry, research has emphasized the importance of user participation and scenario adaptability. Prieto [47] suggested training a GPT model specifically for construction applications, testing it in more complex scenarios, and involving a broader range of participants to improve the generalizability of results. Additionally, Liu [57] proposed integrating knowledge graph feature learning to develop personalized services tailored to user needs. Park [62] further recommended designing customized interactive interfaces, such as voice and gesture control, to enhance acceptance among construction workers. You [63] emphasized that optimizing model decision-making through reinforcement learning from human feedback could improve the applicability and reliability of intelligent systems in construction projects. Nyqvist [53] advocated for applying AI solutions in real-world construction projects to observe their impact and refine project management processes beyond controlled theoretical conditions. More importantly, future research should develop comprehensive evaluation methods to quantify the actual effects of AI tools on project efficiency and safety. This will be crucial for driving widespread adoption and recognition of AI technologies in the construction industry.

5.3.2. Economic Challenges and Efficiency Optimization

The high computational and maintenance costs of deploying generative AI remain a significant barrier to large-scale adoption in the AECO industry, particularly for small- and medium-sized enterprises with limited budgets. Both the training and operational phases of LLMs involve substantial resource consumption. Training demands extensive time and computing power, while real-time interaction requires ongoing energy and server infrastructure to support high-frequency queries [77].

To address these economic constraints, future research should focus on improving system efficiency through lightweight model architectures and query optimization strategies. For instance, in the development of DAVE, a GPT-powered assistant for real-time BIM interaction, researchers observed that each request consumed approximately 3700 tokens in instructions but returned fewer than 100 tokens in response, resulting in high per-request costs [30]. This highlights the importance of reducing unnecessary token consumption and implementing intelligent query management mechanisms to maintain economic viability.

Additionally, future work should explore alternatives to proprietary LLMs—such as open-source models or domain-specific compact models—that can be fine-tuned at lower costs. Federated learning and edge AI strategies may also offer cost-effective deployment options by distributing processing workloads without relying entirely on cloud-based resources. By integrating such cost-reduction methods, generative AI tools can become more accessible across the construction sector, promoting equitable innovation and broader adoption.

5.3.3. Ethics and Explainability

As generative AI becomes more prevalent, its “black-box” problem has emerged as an increasingly pressing challenge. Yoo [60] called for the adoption of saliency visualization techniques to explain the decision-making process of models. For example, visualizing accident prediction logic can help users understand the rationale behind model outputs.

Additionally, enhancing compliance traceability through ontology-based knowledge repositories is an effective way to ensure regulatory adherence [31]. He [44] suggested that integrating domain-specific ontologies could further improve the robustness and generalizability of BERT-based models while mitigating ethical concerns associated with generative AI applications. Moreover, establishing a well-defined framework for data security and accountability is critical to preventing model misuse and ensuring sustainable industry development. For instance, federated learning techniques can address privacy concerns by enabling decentralized model training while maintaining high security and regulatory compliance [54]. By ensuring ethical integrity and explainability, generative AI can exert a long-term and profound impact on the construction industry.

6. Conclusions

The integration of LLMs into the AECO industry signifies a paradigm shift in information processing and knowledge management. This study systematically reviews the current applications and developmental trajectories of LLMs within the AECO sector, highlighting their advancements in handling complex, unstructured data. LLMs have demonstrated exceptional performance in key tasks such as project management, contract analysis, risk assessment, and compliance monitoring. By effectively transforming unstructured data—like construction logs and contractual documents—into structured, actionable insights, LLMs significantly enhance data usability, decision-making efficiency, and process automation, establishing themselves as pivotal tools in the digital transformation of the AEC industry.

However, this study also uncovers several critical challenges and unforeseen limitations. Foremost among these is the interoperability issue between LLMs and prevalent AECO software platforms, characterized by discrepancies in data formats, semantic inconsistencies, restricted API access, and the complexity of proprietary data structures, all of which impede seamless integration and practical application. Additionally, the current capabilities of LLMs in processing multimodal data—encompassing images, videos, and sensor inputs—remain inadequate, falling short of the real-time perception and responsiveness required in dynamic construction environments. While some studies have explored multimodal data fusion to enhance LLM performance, their effectiveness in industrial-scale applications necessitates further validation.

Furthermore, the performance of LLMs in specific applications such as contract risk identification and safety planning for high-risk construction activities is inconsistent. In scenarios involving complex operational contexts—like high-altitude welding or subterranean excavation—the safety recommendations generated by LLMs tend to be overly generic, lacking the specificity needed for actionable guidance. This indicates that, despite their robust natural language processing capabilities, LLMs require enhanced adaptability to specialized engineering contexts.

Future research should concentrate on the following areas: (1) optimizing algorithms and model architectures through improved pretraining strategies, dynamic prompting techniques, and lightweight model designs to enhance reasoning efficiency and adaptability; (2) developing high-quality, multilingual, and multimodal datasets tailored to the unique requirements of the AEC industry to strengthen model comprehension of complex engineering contexts; (3) creating standardized interfaces and embedded intelligent assistants to facilitate deeper integration of LLMs with existing BIM/CAD/IoT platforms; (4) enhancing user engagement mechanisms to improve model applicability and acceptance in complex construction projects; and (5) emphasizing model interpretability and compliance by incorporating knowledge graphs and visualization tools to ensure transparency and reliability in high-risk engineering applications.

Author Contributions

Conceptualization, G.Z. and C.L.; methodology, C.L.; resources, C.L. and Q.L.; writing—original draft preparation, C.L.; writing—review and editing, G.Z. and Q.L.; visualization, C.L.; supervision, Q.L. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

The raw data supporting the conclusions of this article will be made available by the authors on request.

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Table

Figure 1 Research methodology roadmap.

Figure 2 Six-stage systematic review process.

Figure 3 Keywords clustering.

Figure 4 The workflow of the LLMs toolchain.

Figure 5 The workflow of the LLMs toolchain.

Table 1

Table of core technologies and application scenarios of LLMs in the AECO industry.

No.	AECO Phase	Applications	Core Technologies	Papers
1	A—Architecture	Design Assistance and Optimization	Prompt engineering; generative design; natural language interfaces	[8,29,30,31,32,33]
		Cost Estimation of Civil Engineering Works	Domain-specific fine-tuning; information extraction	[34]
		Information Retrieval and Integration	Semantic search; knowledge embedding; retrieval-augmented generation	[11,35,36,37,38,39,40]
2	E—Engineering	Geotechnical Engineering	Multimodal input (text + code); technical document parsing	[41,42]
		Structural Engineering	LLM-to-code generation; parameterized prompting	[43]
3	C—Construction	Construction Management and Optimization	Scheduling LLMs; dialogue-based assistants; task planning	[44,45,46,47]
		Risk Management	Knowledge graph integration; temporal and causal reasoning	[48,49,50,51,52,53]
		Safety Management	Vision-language models; scene graph fusion; hazard recognition	[54,55,56,57,58,59,60,61]
		Construction Automation and Robotics Technology	LLM-to-robot code pipelines; real-time prompt adaptation; human–robot collaboration	[62,63,64,65]
4	O—Operation	Facility Management and Optimization	Decision-support via text-to-action; RAG; semantic enrichment	[66]
		Quality Control and Defect Management	NLP-based defect detection; sentiment analysis for stakeholder feedback	[67]
		Energy Management	Data-to-text generation; energy ontology integration; semantic similarity	[68]
5	Cross-Phase Applications	Knowledge Management and Information Sharing	Retrieval-augmented generation; ontology-grounded LLMs; human-in-the-loop learning	[69,70]
		Contract Management	Legal text mining; clause classification; domain adaptation	[16,71,72,73]
		Construction Activity Monitoring	Video captioning; spatio-temporal reasoning; scene understanding	[74]
		Strategic Technology Integration across AECO Phases	Multiphase AI frameworks; LangChain orchestration; model interoperability	[3,13,75,76,77,78]
		Intelligent Q&A and Interaction	Task-oriented dialogue; semantic parsing; BIM-integrated Q&A interfaces	[79]

References

1. Engineers, B. Bim Handbook: A Guide to Building Information Modeling for Owners Managers; Wiley: Hoboken, NJ, USA, 2018.

2. Succar, B. Building information modelling framework: A research and delivery foundation for industry stakeholders. Autom. Constr.; 2009; 18, pp. 357-375. [DOI: https://dx.doi.org/10.1016/j.autcon.2008.10.003]

3. Wu, C.; Li, X.; Guo, Y.; Wang, J.; Ren, Z.; Wang, M.; Yang, Z. Natural language processing for smart construction: Current status and future directions. Autom. Constr.; 2022; 134, 104059. [DOI: https://dx.doi.org/10.1016/j.autcon.2021.104059]

4. Nasar, Z.; Jaffry, S.W.; Malik, M.K. Named entity recognition and relation extraction: State-of-the-art. ACM Comput. Surv. (CSUR); 2021; 54, pp. 1-39. [DOI: https://dx.doi.org/10.1145/3445965]

5. Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J.D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A. Language models are few-shot learners. Adv. Neural Inf. Process. Syst.; 2020; 33, pp. 1877-1901.

6. Hadi, M.U.; Qureshi, R.; Shah, A.; Irfan, M.; Zafar, A.; Shaikh, M.B.; Akhtar, N.; Wu, J.; Mirjalili, S. Large language models: A comprehensive survey of its applications, challenges, limitations, and future prospects. Authorea Prepr.; 2023; 1, pp. 1-26.

7. Wang, Z.; Li, Y.; Zhu, Z. BERT-based knowledge extraction method of unstructured domain text. arXiv; 2021; arXiv: 2103.00728

8. Jiang, C.; Zheng, Z.; Liang, X.; Lin, J.; Ma, Z.; Lu, X. A new interaction paradigm for building design driven by large language model: Proof of concept with Rhino7. J. Graph.; 2024; 45, 594.

9. Zhao, W.X.; Zhou, K.; Li, J.; Tang, T.; Wang, X.; Hou, Y.; Min, Y.; Zhang, B.; Zhang, J.; Dong, Z. A survey of large language models. arXiv; 2023; arXiv: 2303.18223

10. Chang, Y.; Wang, X.; Wang, J.; Wu, Y.; Yang, L.; Zhu, K.; Chen, H.; Yi, X.; Wang, C.; Wang, Y. A survey on evaluation of large language models. ACM Trans. Intell. Syst. Technol.; 2024; 15, pp. 1-45. [DOI: https://dx.doi.org/10.1145/3641289]

11. Shamshiri, A.; Ryu, K.R.; Park, J.Y. Text mining and natural language processing in construction. Autom. Constr.; 2024; 158, 105200. [DOI: https://dx.doi.org/10.1016/j.autcon.2023.105200]

12. Machi, L.A.; McEvoy, B.T. The Literature Review: Six Steps to Success; SAGE Publications: Thousand Oaks, CA, USA, 2009.

13. Onatayo, D.; Onososen, A.; Oyediran, A.O.; Oyediran, H.; Arowoiya, V.; Onatayo, E. Generative AI Applications in Architecture, Engineering, and Construction: Trends, Implications for Practice, Education & Imperatives for Upskilling—A Review. Architecture; 2024; 4, pp. 877-902. [DOI: https://dx.doi.org/10.3390/architecture4040046]

14. Onatayo, D.A.; Srinivasan, R.S.; Shah, B. Ultraviolet Radiation Transmission in Buildings’ Fenestration: Part I, Detection Methods and Approaches Using Spectrophotometer and Radiometer. Buildings; 2023; 13, 1670. [DOI: https://dx.doi.org/10.3390/buildings13071670]

15. Raiaan, M.A.K.; Mukta, M.S.H.; Fatema, K.; Fahad, N.M.; Sakib, S.; Mim, M.M.J.; Ahmad, J.; Ali, M.E.; Azam, S. A Review on Large Language Models: Architectures, Applications, Taxonomies, Open Issues and Challenges. IEEE Access; 2024; 12, pp. 26839-26874. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3365742]

16. Jafari, P.; Al Hattab, M.; Mohamed, E.; AbouRizk, S. Automated extraction and time-cost prediction of contractual reporting requirements in construction using natural language processing and simulation. Appl. Sci.; 2021; 11, 6188. [DOI: https://dx.doi.org/10.3390/app11136188]

17. Raschka, S. Build a Large Language Model (From Scratch); Simon and Schuster: New York, NY, USA, 2024.

18. Soori, M.; Jough, F.K.G.; Dastres, R.; Arezoo, B. AI-based decision support systems in Industry 4.0, A review. J. Econ. Technol.; 2024; in press

19. Taye, M.M. Understanding of machine learning with deep learning: Architectures, workflow, applications and future directions. Computers; 2023; 12, 91. [DOI: https://dx.doi.org/10.3390/computers12050091]

20. Najafabadi, M.M.; Villanustre, F.; Khoshgoftaar, T.M.; Seliya, N.; Wald, R.; Muharemagic, E. Deep learning applications and challenges in big data analytics. J. Big Data; 2015; 2, 1. [DOI: https://dx.doi.org/10.1186/s40537-014-0007-7]

21. Mohy, A.A.; Bassioni, H.A.; Elgendi, E.O.; Hassan, T.M. Innovations in safety management for construction sites: The role of deep learning and computer vision techniques. Constr. Innov.; 2024; ahead of print

22. Bonyani, M.; Soleymani, M. Towards improving workers’ safety and progress monitoring of construction sites through construction site understanding. arXiv; 2022; arXiv: 2210.15760

23. Akinosho, T.D.; Oyedele, L.O.; Bilal, M.; Ajayi, A.O.; Delgado, M.D.; Akinade, O.O.; Ahmed, A.A. Deep learning in the construction industry: A review of present status and future innovations. J. Build. Eng.; 2020; 32, 101827. [DOI: https://dx.doi.org/10.1016/j.jobe.2020.101827]

24. Chai, P.; Hou, L.; Zhang, G.; Tushar, Q.; Zou, Y. Generative adversarial networks in construction applications. Autom. Constr.; 2024; 159, 105265. [DOI: https://dx.doi.org/10.1016/j.autcon.2024.105265]

25. Regenwetter, L.; Nobari, A.H.; Ahmed, F. Deep generative models in engineering design: A review. J. Mech. Des.; 2022; 144, 071704. [DOI: https://dx.doi.org/10.1115/1.4053859]

26. Lee, D.; Lee, S.; Masoud, N.; Krishnan, M.; Li, V.C. Digital twin-driven deep reinforcement learning for adaptive task allocation in robotic construction. Adv. Eng. Inform.; 2022; 53, 101710. [DOI: https://dx.doi.org/10.1016/j.aei.2022.101710]

27. Jia, Y.; Wang, J.; Shou, W.; Hosseini, M.R.; Bai, Y. Graph neural networks for construction applications. Autom. Constr.; 2023; 154, 104984. [DOI: https://dx.doi.org/10.1016/j.autcon.2023.104984]

28. Chen, G.; Hou, J.; Liu, C.; Hu, K.; Wang, J. Visualization analysis of cross research between big data and construction industry based on knowledge graph. Buildings; 2022; 12, 1812. [DOI: https://dx.doi.org/10.3390/buildings12111812]

29. Zheng, J.; Fischer, M. Dynamic prompt-based virtual assistant framework for BIM information search. Autom. Constr.; 2023; 155, 105067. [DOI: https://dx.doi.org/10.1016/j.autcon.2023.105067]

30. Fernandes, D.; Garg, S.; Nikkel, M.; Guven, G. A GPT-Powered Assistant for Real-Time Interaction with Building Information Models. Buildings; 2024; 14, 2499. [DOI: https://dx.doi.org/10.3390/buildings14082499]

31. Chen, N.; Lin, X.; Jiang, H.; An, Y. Automated Building Information Modeling Compliance Check through a Large Language Model Combined with Deep Learning and Ontology. Buildings; 2024; 14, 1983. [DOI: https://dx.doi.org/10.3390/buildings14071983]

32. Dortheimer, J.; Martelaro, N.; Sprecher, A.; Schubert, G. Evaluating large-language-model chatbots to engage communities in large-scale design projects. Artif. Intell. Eng. Des. Anal. Manuf.; 2024; 38, e4. [DOI: https://dx.doi.org/10.1017/S0890060424000027]

33. Forth, K.; Borrmann, A. Semantic enrichment for BIM-based building energy performance simulations using semantic textual similarity and fine-tuning multilingual LLM. J. Build. Eng.; 2024; 95, 110312. [DOI: https://dx.doi.org/10.1016/j.jobe.2024.110312]

34. Kang, T.; Kang, K. Earthwork Network Architecture (ENA): Research for Earthwork Quantity Estimation Method Improvement with Large Language Model. Appl. Sci.; 2024; 14, 10517. [DOI: https://dx.doi.org/10.3390/app142210517]

35. Al-Turki, D.; Hettiarachchi, H.; Gaber, M.M.; Abdelsamea, M.M.; Basurra, S.; Iranmanesh, S.; Saadany, H.; Vakaj, E. Human-in-the-Loop Learning with LLMs for Efficient RASE Tagging in Building Compliance Regulations. IEEE Access; 2024; 12, pp. 185291-185306. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3512434]

36. Zhong, Y.; Goodfellow, S.D. Domain-specific language models pre-trained on construction management systems corpora. Autom. Constr.; 2024; 160, 105316. [DOI: https://dx.doi.org/10.1016/j.autcon.2024.105316]

37. Zheng, Z.; Lu, X.-Z.; Chen, K.-Y.; Zhou, Y.-C.; Lin, J.-R. Pretrained domain-specific language model for general information retrieval tasks in the AEC domain. arXiv; 2022; arXiv: 2203.04729

38. Gao, B.; Hu, Y.; Gu, J.; Han, X. Integrating deep learning and multi-attention for joint extraction of entities and relationships in engineering consulting texts. Autom. Constr.; 2024; 168, 105739. [DOI: https://dx.doi.org/10.1016/j.autcon.2024.105739]

39. Li, H.; Yang, R.; Xu, S.; Xiao, Y.; Zhao, H. Intelligent Checking Method for Construction Schemes via Fusion of Knowledge Graph and Large Language Models. Buildings; 2024; 14, 2502. [DOI: https://dx.doi.org/10.3390/buildings14082502]

40. Lee, J.; Jung, W.; Baek, S. In-House Knowledge Management Using a Large Language Model: Focusing on Technical Specification Documents Review. Appl. Sci.; 2024; 14, 2096. [DOI: https://dx.doi.org/10.3390/app14052096]

41. Kim, D.; Kim, T.; Kim, Y.; Byun, Y.-H.; Yun, T.S. A ChatGPT-MATLAB framework for numerical modeling in geotechnical engineering applications. Comput. Geotech.; 2024; 169, 106237. [DOI: https://dx.doi.org/10.1016/j.compgeo.2024.106237]

42. Kumar, K. Geotechnical parrot tales (gpt): Harnessing large language models in geotechnical engineering. J. Geotech. Geoenvironmental Eng.; 2024; 150, 02523001. [DOI: https://dx.doi.org/10.1061/JGGEFK.GTENG-11828]

43. Tian, S.; Jiang, X.; Wang, W.; Jing, Z.; Zhang, C.; Zhang, C.; Lookman, T.; Su, Y. Steel design based on a large language model. Acta Mater.; 2025; 285, 120663. [DOI: https://dx.doi.org/10.1016/j.actamat.2024.120663]

44. He, C.; Yu, B.; Liu, M.; Guo, L.; Tian, L.; Huang, J. Utilizing Large Language Models to Illustrate Constraints for Construction Planning. Buildings; 2024; 14, 2511. [DOI: https://dx.doi.org/10.3390/buildings14082511]

45. Chen, G.; Alsharef, A.; Ovid, A.; Albert, A.; Jaselskis, E. Meet2Mitigate: An LLM-powered framework for real-time issue identification and mitigation from construction meeting discourse. Adv. Eng. Inform.; 2025; 64, 103068. [DOI: https://dx.doi.org/10.1016/j.aei.2024.103068]

46. Bah Esseme, A.; Faniyan, A.; Farayola, G. AI in Project Management: Enhancing Efficiency. Decis. Mak. Risk Manag. J. Artif. Intell. Mach. Learn. Data Sci.; 2025; 3, pp. 2107-2114.

47. Prieto, S.A.; Mengiste, E.T.; García de Soto, B. Investigating the Use of ChatGPT for the Scheduling of Construction Projects. Buildings; 2023; 13, 857. [DOI: https://dx.doi.org/10.3390/buildings13040857]

48. Shishehgarkhaneh, M.B.; Moehler, R.C.; Fang, Y.; Hijazi, A.A.; Aboutorab, H. Transformer-Based Named Entity Recognition in Construction Supply Chain Risk Management in Australia. IEEE Access; 2024; 12, pp. 41829-41851. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3377232]

49. Zhang, L.; Wang, J.; Wang, Y.; Sun, H.; Zhao, X. Automatic construction site hazard identification integrating construction scene graphs with BERT based domain knowledge. Autom. Constr.; 2022; 142, 104535. [DOI: https://dx.doi.org/10.1016/j.autcon.2022.104535]

50. Lee, W.; Lee, S. Development of a Knowledge Base for Construction Risk Assessments Using BERT and Graph Models. Buildings; 2024; 14, 3359. [DOI: https://dx.doi.org/10.3390/buildings14113359]

51. Zhou, H.; Tang, S.; Huang, W.; Zhao, X. Generating risk response measures for subway construction by fusion of knowledge and deep learning. Autom. Constr.; 2023; 152, 104951. [DOI: https://dx.doi.org/10.1016/j.autcon.2023.104951]

52. Cao, K.; Chen, S.; Yang, C.; Li, Z.; Luo, L.; Ren, Z. Revealing the coupled evolution process of construction risks in mega hydropower engineering through textual semantics. Adv. Eng. Inform.; 2024; 62, 102713. [DOI: https://dx.doi.org/10.1016/j.aei.2024.102713]

53. Nyqvist, R.; Peltokorpi, A.; Seppänen, O. Can ChatGPT exceed humans in construction project risk management?. Eng. Constr. Archit. Manag.; 2024; 31, pp. 223-243. [DOI: https://dx.doi.org/10.1108/ECAM-08-2023-0819]

54. Zhong, B.; Shen, L.; Pan, X.; Lei, L. Visual attention framework for identifying semantic information from construction monitoring video. Saf. Sci.; 2023; 163, 106122. [DOI: https://dx.doi.org/10.1016/j.ssci.2023.106122]

55. Shin, S.-H.; Won, J.-H.; Jeong, H.-J.; Kang, M.-G. Development of a Site Information Classification Model and a Similar-Site Accident Retrieval Model for Construction Using the KLUE-BERT Model. Buildings; 2024; 14, 1797. [DOI: https://dx.doi.org/10.3390/buildings14061797]

56. Tian, D.; Li, M.; Ren, Q.; Zhang, X.; Han, S.; Shen, Y. Intelligent question answering method for construction safety hazard knowledge based on deep semantic mining. Autom. Constr.; 2023; 145, 104670. [DOI: https://dx.doi.org/10.1016/j.autcon.2022.104670]

57. Liu, X.; Jing, X.; Zhu, Q.; Du, W.; Wang, X. Automatic construction hazard Identification integrating on-site scene graphs with information extraction in outfield test. Buildings; 2023; 13, 377. [DOI: https://dx.doi.org/10.3390/buildings13020377]

58. Smetana, M.; Salles de Salles, L.; Sukharev, I.; Khazanovich, L. Highway Construction Safety Analysis Using Large Language Models. Appl. Sci.; 2024; 14, 1352. [DOI: https://dx.doi.org/10.3390/app14041352]

59. Uhm, M.; Kim, J.; Ahn, S.; Jeong, H.; Kim, H. Effectiveness of retrieval augmented generation-based large language models for generating construction safety information. Autom. Constr.; 2025; 170, 105926. [DOI: https://dx.doi.org/10.1016/j.autcon.2024.105926]

60. Yoo, B.; Kim, J.; Park, S.; Ahn, C.R.; Oh, T. Harnessing Generative Pre-Trained Transformers for Construction Accident Prediction with Saliency Visualization. Appl. Sci.; 2024; 14, 664. [DOI: https://dx.doi.org/10.3390/app14020664]

61. Lee, J.; Ahn, S.; Kim, D.; Kim, D. Performance comparison of retrieval-augmented generation and fine-tuned large language models for construction safety management knowledge retrieval. Autom. Constr.; 2024; 168, 105846. [DOI: https://dx.doi.org/10.1016/j.autcon.2024.105846]

62. Park, S.; Menassa, C.C.; Kamat, V.R. Integrating large language models with multimodal virtual reality interfaces to support collaborative human–robot construction work. J. Comput. Civ. Eng.; 2025; 39, 04024053. [DOI: https://dx.doi.org/10.1061/JCCEE5.CPENG-6106]

63. You, H.; Ye, Y.; Zhou, T.; Zhu, Q.; Du, J. Robot-Enabled Construction Assembly with Automated Sequence Planning Based on ChatGPT: RoboGPT. Buildings; 2023; 13, 1772. [DOI: https://dx.doi.org/10.3390/buildings13071772]

64. Pu, H.; Yang, X.; Li, J.; Guo, R. AutoRepo: A general framework for multimodal LLM-based automated construction reporting. Expert Syst. Appl.; 2024; 255, 124601. [DOI: https://dx.doi.org/10.1016/j.eswa.2024.124601]

65. Luo, H.; Wu, J.; Liu, J.; Antwi-Afari, M.F. Large language model-based code generation for the control of construction assembly robots: A hierarchical generation approach. Dev. Built Environ.; 2024; 19, 100488. [DOI: https://dx.doi.org/10.1016/j.dibe.2024.100488]

66. Zhang, C.; Lei, X.; Xia, Y.; Sun, L. Automatic bridge inspection database construction through hybrid information extraction and large language models. Dev. Built Environ.; 2024; 20, 100549. [DOI: https://dx.doi.org/10.1016/j.dibe.2024.100549]

67. Shooshtarian, S.; Gurmu, A.T.; Sadick, A.-M. Application of natural language processing in residential building defects analysis: Australian stakeholders’ perceptions, causes and types. Eng. Appl. Artif. Intell.; 2023; 126, 107178. [DOI: https://dx.doi.org/10.1016/j.engappai.2023.107178]

68. Liu, M.; Zhang, L.; Chen, J.; Chen, W.-A.; Yang, Z.; Lo, L.J.; Wen, J.; O’Neill, Z. Large language models for building energy applications: Opportunities and challenges. Build. Simul.; 2025; 18, pp. 225-234. [DOI: https://dx.doi.org/10.1007/s12273-025-1235-9]

69. Xu, F.; Nguyen, T.; Du, J. Augmented Reality for Maintenance Tasks with ChatGPT for Automated Text-to-Action. J. Constr. Eng. Manag.; 2024; 150, 04024015. [DOI: https://dx.doi.org/10.1061/JCEMD4.COENG-14142]

70. Yu, Z.; Fung, B.C.; Haghighat, F. Extracting knowledge from building-related data—A data mining framework. Proceedings of the Building Simulation; Chambery, France, 25–28 August 2013; pp. 207-222.

71. Moon, S.; Chi, S.; Im, S.-B. Automated detection of contractual risk clauses from construction specifications using bidirectional encoder representations from transformers (BERT). Autom. Constr.; 2022; 142, 104465. [DOI: https://dx.doi.org/10.1016/j.autcon.2022.104465]

72. Wong, S.; Zheng, C.; Su, X.; Tang, Y. Construction contract risk identification based on knowledge-augmented language models. Comput. Ind.; 2024; 157–158, 104082. [DOI: https://dx.doi.org/10.1016/j.compind.2024.104082]

73. Kazemi, M.H.; Alvanchi, A. Application of NLP-based models in automated detection of risky contract statements written in complex script system. Expert Syst. Appl.; 2025; 259, 125296. [DOI: https://dx.doi.org/10.1016/j.eswa.2024.125296]

74. Pan, Z.; Su, C.; Deng, Y.; Cheng, J. Image2Triplets: A computer vision-based explicit relationship extraction framework for updating construction activity knowledge graphs. Comput. Ind.; 2022; 137, 103610. [DOI: https://dx.doi.org/10.1016/j.compind.2022.103610]

75. Saka, A.; Taiwo, R.; Saka, N.; Salami, B.A.; Ajayi, S.; Akande, K.; Kazemi, H. GPT models in construction industry: Opportunities, limitations, and a use case validation. Dev. Built Environ.; 2024; 17, 100300. [DOI: https://dx.doi.org/10.1016/j.dibe.2023.100300]

76. Rane, N.; Choudhary, S.; Rane, J. Integrating ChatGPT, Bard, and Leading-Edge Generative Artificial Intelligence in Building and Construction Industry: Applications, Framework, Challenges, and Future Scope. 2023; Available online: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4645597 (accessed on 16 April 2025).

77. Ghimire, P.; Kim, K.; Acharya, M. Opportunities and Challenges of Generative AI in Construction Industry: Focusing on Adoption of Text-Based Models. Buildings; 2024; 14, 220. [DOI: https://dx.doi.org/10.3390/buildings14010220]

78. Taiwo, R.; Bello, I.T.; Abdulai, S.F.; Yussif, A.-M.; Salami, B.A.; Saka, A.; Seghier, M.E.A.B.; Zayed, T. Generative artificial intelligence in construction: A Delphi approach, framework, and case study. Alex. Eng. J.; 2025; 116, pp. 672-698. [DOI: https://dx.doi.org/10.1016/j.aej.2024.12.079]

79. Lin, T.-H.; Huang, Y.-H.; Putranto, A. Intelligent question and answer system for building information modeling and artificial intelligence of things based on the bidirectional encoder representations from transformers model. Autom. Constr.; 2022; 142, 104483. [DOI: https://dx.doi.org/10.1016/j.autcon.2022.104483]

80. Lin, J.R.; Hu, Z.Z.; Zhang, J.P.; Yu, F.Q. A natural-language-based approach to intelligent data retrieval and representation for cloud BIM. Comput.-Aided Civ. Infrastruct. Eng.; 2016; 31, pp. 18-33. [DOI: https://dx.doi.org/10.1111/mice.12151]

81. Maalek, R. Integrating Generative Artificial Intelligence and Problem-Based Learning into the Digitization in Construction Curriculum. Buildings; 2024; 14, 3642. [DOI: https://dx.doi.org/10.3390/buildings14113642]

82. Ngo, T.T.A.; An, G.K.; Nguyen, P.T.; Tran, T.T. Unlocking educational potential: Exploring students’ satisfaction and sustainable engagement with ChatGPT using the ECM model. J. Inf. Technol. Educ. Res.; 2024; 23, 021.

83. Jeong, J.; Gil, D.; Kim, D.; Jeong, J. Current Research and Future Directions for Off-Site Construction through LangChain with a Large Language Model. Buildings; 2024; 14, 2374. [DOI: https://dx.doi.org/10.3390/buildings14082374]

84. Yang, Y.; Wu, Z.; Yang, Y.; Lian, S.; Guo, F.; Wang, Z. A Survey of Information Extraction Based on Deep Learning. Appl. Sci.; 2022; 12, 9691. [DOI: https://dx.doi.org/10.3390/app12199691]

85. Pu, H.; Chen, W. Review of multimodal named entity recognition studies. Data Anal. Knowl. Discov.; 2024; 8, pp. 50-63.

86. Swarup, A.; Bhandarkar, A.; Dizon-Paradis, O.P.; Wilson, R.; Woodard, D.L. Maximizing Relation Extraction Potential: A Data-Centric Study to Unveil Challenges and Opportunities. IEEE Access; 2024; 12, pp. 167655-167682. [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3494737]

87. Tang, J.; Song, S.; Du, Q.; Yao, Y.; Qu, J. Graph convolutional networks with the self-attention mechanism for adaptive influence maximization in social networks. Complex Intell. Syst.; 2024; 10, pp. 8383-8401. [DOI: https://dx.doi.org/10.1007/s40747-024-01604-y]

88. Guo, Z.; Zhang, Y.; Lu, W. Attention guided graph convolutional networks for relation extraction. arXiv; 2019; arXiv: 1906.07510

89. Mirza, P.; Tonelli, S. Catena: Causal and temporal relation extraction from natural language texts. Proceedings of the 26th International Conference on Computational Linguistics; Osaka, Japan, 11–16 December 2016; pp. 64-75.

90. Piskorski, J.; Yangarber, R. Information extraction: Past, present and future. Multi-Source, Multilingual Information Extraction and Summarization; Springer: Berlin/Heidelberg, Germany, 2013; pp. 23-49.

91. Choi, S.; Jung, Y. Knowledge Graph Construction: Extraction, Learning, and Evaluation. Appl. Sci.; 2025; 15, 3727. [DOI: https://dx.doi.org/10.3390/app15073727]

92. Walker, A. Project Management in Construction; John Wiley & Sons: Hoboken, NJ, USA, 2015.

93. Akbarpour, M. Network Theory in Engineering Project Management: A Review of Graph-Based Models for Optimization. Manag. Strateg. Eng. Sci.; 2023; 5, pp. 32-41.

94. Venugopal, M.; Eastman, C.M.; Sacks, R.; Teizer, J. Semantics of model views for information exchanges using the industry foundation class schema. Adv. Eng. Inform.; 2012; 26, pp. 411-428. [DOI: https://dx.doi.org/10.1016/j.aei.2012.01.005]

95. Paskoff, C.; Boton, C.; Blanchet, P. Comparative analyses of four BIM-based compliance checking tools. Proceedings of the Canadian Society of Civil Engineering Annual Conference; Moncton, NB, Canada, 24–27 May 2023; pp. 461-471.

96.ISO 19650-1: 2018 BSI Standards Publication Organization and Digitization of Information About Buildings and Civil Engineering Works, Including Building Information Modelling (BIM)-Information Management Using Building Information Modelling; ISO: Geneva, Switzerland, 2018.

97. Gao, Y.; Xiong, Y.; Gao, X.; Jia, K.; Pan, J.; Bi, Y.; Dai, Y.; Sun, J.; Wang, H.; Wang, H. Retrieval-augmented generation for large language models: A survey. arXiv; 2023; arXiv: 2312.10997

98. Oh, T.K.; Kwon, Y.J.; Oh, B.-H.; Gwon, Y.-I.; Yoon, H.-K. Suggestions for safety coordinator’s roles at each construction stage (client, designer, supervisor, and contractor) to improve safety and health activities in South Korea. Saf. Sci.; 2021; 133, 104994.

99. Guidotti, R.; Monreale, A.; Ruggieri, S.; Turini, F.; Giannotti, F.; Pedreschi, D. A survey of methods for explaining black box models. ACM Comput. Surv. (CSUR); 2018; 51, pp. 1-42. [DOI: https://dx.doi.org/10.1145/3236009]

100. Wu, B.; Liu, S.; Zhong, Y.; Sun, P.; Liu, X.; Jin, X. LoongServe: Efficiently Serving Long-Context Large Language Models with Elastic Sequence Parallelism. Proceedings of the ACM SIGOPS 30th Symposium on Operating Systems Principles; Austin, TX, USA, 4–6 November 2024; pp. 640-654.