Developments in The Built Environment

Developments in the Built Environment 17 (2024) 100300
Contents lists available at ScienceDirect
Developments in the Built Environment

journal homepage: www.sciencedirect.com/journal/developments-in-the-built-environment
GPT models in construction industry: Opportunities, limitations, and a use

case validation
Abdullahi Saka a, Ridwan Taiwo b, Nurudeen Saka a, Babatunde Abiodun Salami c,
Saheed Ajayi a, *, Kabiru Akande d, Hadi Kazemi a
a
School of Built Environment, Engineering and Computing, Leeds Beckett University, UK
b
Department of Building and Real Estate, Hong Kong Polytechnic University, Hong Kong
c
Cardiff School of Management, Cardiff Metropolitan University (Llandaff Campus), UK
d
OVO Energy, UK
A R T I C L E I N F O A B S T R A C T
Keywords: Large Language Models (LLMs) trained on large data sets came into prominence in 2018 after Google introduced
LLMs BERT. Subsequently, different LLMs such as GPT models from OpenAI have been released. These models perform
ChatGPT well on diverse tasks and have been gaining widespread applications in fields such as business and education.
GPT
However, little is known about the opportunities and challenges of using LLMs in the construction industry. Thus,
Artificial intelligence
Generative AI
this study aims to assess GPT models in the construction industry. A critical review, expert discussion and case
study validation are employed to achieve the study’s objectives. The findings revealed opportunities for GPT
models throughout the project lifecycle. The challenges of leveraging GPT models are highlighted and a use case
prototype is developed for materials selection and optimization. The findings of the study would be of benefit to
researchers, practitioners and stakeholders, as it presents research vistas for LLMs in the construction industry.
1. Introduction advancement in information technologies and digital tools, the AEC

industry has been embracing its usage to improve its performance in a
The architecture, engineering, and construction (AEC) industry is bid towards the fourth industrial revolution (industry 4.0). Conse
known for its slow adoption of innovation, when compared to other quently, there has been an increase in the usage of building information
industries, due to the culture of the industry and the nature of its modelling (BIM), application of big data analytics, offsite construction,
products (Gambatese and Hallowell, 2011). The industry is automation, and artificial intelligence (AI).
information-intensive and relies on myriad and diverse information AI deals with the ability of machines to perform tasks that typically
from different stakeholders for successful project delivery (Chen and require human intelligence, such as learning, reasoning, perception and
Kamara, 2005). However, there is a lack of information integration, decision-making. AI systems process and analyse large datasets with the
reuse, and efficient management, all of which have a tremendous effect view of identifying patterns, relationships, drawing inferences, recom
on stakeholders’ collaboration and productivity of the industry. Past mendations and taking action. Abioye et al. (2021) listed the subfields of
reports in the construction industry have emphasized the need for AI to include machine learning, knowledge-based systems, computer
improvement in the modus Operandi of the industry to improve pro vision, robotics, natural language processing, automated planning and
ductivity and achieve value for money (Egan, 1998). Although the in scheduling and optimization. These diverse fields have been employed
dustry currently contributes about 13% to the global GDP, productivity in the AEC industry to improve productivity and efficiency. As such, AI
growth has only been increasing at 1% per year over the last two decades has been leveraged in cost prediction, delay prediction, building design
(Ribeirinho et al., 2020). Also, the industry is facing a myriad of chal energy prediction, workers’ activity recognition, construction site
lenges such as delays, health & safety, cost overrun, shortage of skilled safety, cash flow prediction, structural health monitoring, resource
personnel, and stringent requirements by governments. With the allocation and optimization, predictive maintenance, and decision
* Corresponding author.
E-mail addresses: [email protected] (A. Saka), [email protected] (R. Taiwo), [email protected] (N. Saka), basalami@
cardiffmet.ac.uk (B.A. Salami), [email protected] (S. Ajayi), [email protected] (K. Akande), [email protected] (H. Kazemi).
https://doi.org/10.1016/j.dibe.2023.100300
Received 30 May 2023; Received in revised form 26 October 2023; Accepted 4 December 2023
Available online 12 December 2023
2666-1659/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
A. Saka et al. Developments in the Built Environment 17 (2024) 100300
support system, among others (Abioye et al., 2021). Studies have shown current study aims to critically review GPT models in the AEC industry
significant improvement in productivity and efficiency with the use of with the following objectives:
these tools; however, bottlenecks have been reported. Challenges such
as lack of skilled workers, cultural resistance to change, cost of imple a) To identify opportunities for the application of GPT models
mentation, unavailability of structured data, trust and ethics have been b) To evaluate the limitations to the application of GPT models in the
highlighted as the major hurdles towards the effective deployment of AI AEC industry
in the AEC industry (Akinosho et al., 2020). c) To validate a use case for GPT models in the AEC industry
With the advancement in AI, there has been significant improvement
in the development of Conversational AI (CAI) and Generative AI (GAI). Achieving these objectives would significantly contribute to the
Conversational AI deals with the application of NLP to enable computers emerging body of knowledge on LLMs and Generative AI in the con
to understand and interact with humans in a conversational way using struction industry and provide research agenda for researchers. Also,
natural language (Kulkarni et al., 2019). Generative AI deals with the this study highlights areas that would benefit from the application of
creation of novel content such as texts, audio, and images (Gozalo- GPT models and provides a case study validation for the use of GPT
Brizuela and Garrido-Merchan, 2023). CAI improved human-computer model in material selection and optimization. Similarly, inherent chal
interactions and led to the development of chatbots, virtual assistants lenges of the application of GPT are presented to enlighten stakeholders
and other conversational interfaces which often leverage GAI and can about the possible pitfalls that can be encountered in the deployment of
assist users in automating tasks, information retrieval, and customer GPT models and to prevent health, safety, and business problems. The
services among others. Saka et al. (2023) reviewed the current appli rest of the paper is structured into seven sections (Fig. 1) covering the
cations of Conversational AI in the AEC industry and highlighted that literature review, methodology, findings, discussion, case validation and
the deployment of this emerging field is still limited. Few extant studies conclusion.
on the application of Conversational AI in the construction industry
focused on information retrieval from BIM with limited functionalities. 2. Generative Pre-trained transformer (GPT)
Majority of the developed Conversational AI agents in the AEC industry
are based on the traditional approach to NLP, which requires time for OpenAI’s Generative Pre-trained (GPT) models have made signifi
processing the data, and users’ interactions are often restricted as the cant contributions to the field of language generation. GPT models use
agents are developed with the assumption of happy path users. Simi transformer-based models that learn statistical patterns of natural lan
larly, other studies have leveraged on machine learning for the devel guage, enabling them to generate human-like language. The series
opment of Conversational agents such as Bidirectional Encoder started with GPT-1 in June 2018 and has since evolved to GPT-2, GPT-3,
Representations from Transformers (BERT) and the use of commercial and GPT-3.5 (OpenAI, 2019; Radford et al., 2019). The latest addition,
platforms such as IBM Watson, Amazon Alexa, Google Natural Language GPT-4, was launched in March 2023 and demonstrates significant ad
AI, and Microsoft Azure (Saka et al., 2023). However, this approach vancements in generating coherent and understandable text. GPT
often requires large data sets for training which are unavailable and models are trained using vast amounts of unstructured text data,
expensive to gather in the AEC industry. enabling them to generate language almost indistinguishable from
Furthermore, these machine learning approaches such as BERT are human-generated text (OpenAI, 2023a,b). Early NLP relied on
part of LLMs which came to the limelight in 2018 after the introduction rule-based systems that required explicit programming of grammar rules
of transformer-a model architecture which rely on attention mechanism and syntax. However, these systems have limitations in programming
and differs from recurrent neural networks - by Vaswani et al. (2017). complex languages and linguistic nuances, making them less adaptable
LLMs are neural networks with large parameters and are trained using to new domains or contexts and less scalable (Shaalan, 2010). The rise of
self-supervised learning and semi-supervised learning on large datasets. data-driven approaches, such as GPT models, enabled machine learning
These LLMs have improved NLP and shifted the direction away from algorithms to learn from large amounts of data and recognize complex
training with labelled data for defined objectives. Generative patterns in natural language without explicit programming of rules. The
Pre-trained Transformer (GPT) models which are decoder blocks only GPT-3 API was introduced in June 2020 and made publicly available in
from OpenAI which have gained significant attention and showed November 2021 (Karhade, 2022; OpenAI, 2020). It brought significant
improved performance from GPT-2 (trained with 10 billion tokens) to advancements in NLP technology making GPT-3 widely accessible. In
GPT-3 (trained with 499 billion tokens) and recently GPT-4 released in January 2022, InstructGPT, a version of GPT 3.5, which can handle
2023. The GPT models as a result of the large training dataset and large more complicated instructions, was released. In 2022, speech recogni
parameters have enabled few-shot (provide contexts and examples in tion software Whisper and GPT-3.5 upgrade to text-davinci-003 were
the prompts), zero-shot (no example is provided in the prompt) learning introduced in September and November, respectively (Karhade, 2022;
capability (Wei et al., 2022). As a result, it has been widely deployed in OpenAI, 2023a,b; Radford et al., 2022). GPT-4 has further advanced the
many applications. GPT models have beneficial applications in health NLP after launching in March 2023, opening up new possibilities for
care for triaging, analysing electronic health records, translation, med industry-specific applications. The GPT models have transformed the
ical education, medical and diagnostic (Li et al., 2023). In field of NLP, enabling previously unattainable levels of fluency and
bioinformatics, GPT model have been applied in sequence analysis, coherence in machine-generated text (OpenAI, 2023a,b). Table 1 pro
Genome analysis, Gene expression, proteomics, and in drug discovery vides a chronological summary of significant milestones in the devel
(Zhang et al., 2023). In education, GPT could transform autodidactic opment and release of GPT models.
experience by providing personalized support, increased accessibility, One of the main advantages of GPT models is their capacity to pro
flexible learning, real-time feedback and guidance (Firat, 2023). Simi duce language that is cohesive, fluent, and nearly indistinguishable from
larly, GPT can be leveraged in business and commerce for chatbots, text produced by humans. These models have been effectively used in a
virtual assistants, customer service management, sentiment analysis, variety of applications, including chatbots, content generation, and
financial analysis and forecasting, fraud detection and supply chain machine translation. They can produce answers to open-ended ques
management (Zong and Krishnamachari, 2022). tions, making them an important tool for natural language communi
Despite GPT models overcoming some of the extant challenges of cation. The layers of GPTs’ transformer-based neural architecture
developing AI applications in the construction industry and providing employ attention techniques to concentrate on particular areas of the
opportunities to improve productivity, there are few studies on GPT input text (Neelakantan et al., 2022; Vaswani et al., 2017). The model
models in the AEC industry. Also, there are no reviews on the oppor can pick up on statistical patterns in natural language attributable to its
tunities of these emerging LLMs in the literature. Consequently, this architecture without having to explicitly program it with syntax or
2
Fig. 1. Structure of the paper.
Table 1 Table 3
The progression of OpenAI’s GPT models from GPT-1 to GPT-4. Experts’ demographic details.
Date Milestone References Designation Professional Experience Sector and Expertise
Background
June 11th, OpenAI introduced GPT-1, the first model in MUO (2023)
2018 the GPT series. A Architect 15 years Research and Development with
Feb 14th, OpenAI announced the release of GPT-2. Radford et al. (2019) expertise in the deployment of
2019 Artificial Intelligence
May 28th, OpenAI published the initial GPT-3 preprint Brown et al. (2020) B Software 14 years IT with expertise in developing
2020 paper on arXiv. Developer solutions for AEC companies
June 11th, OpenAI launched a private beta for the GPT- VentureBeat (2020) C Civil Engineer 12 years Research and Development with
2020 3 API. expertise in the deployment of
Sep 22nd, OpenAI licensed GPT-3 to Microsoft. OpenAI (2020) Artificial Intelligence
2020 D Project 14 years Research and Development with
Nov 18th, The GPT-3 API was opened to the public. VentureBeat (2020) Manager expertise in construction project
2021 analytics
Jan 27th, OpenAI released InstructGPT as text-davinci- OpenAI (2023) E Computer 19 years Research and Development with
2022 002, later renamed GPT-3.5. Engineer expertise in the deployment of
Jult 28th OpenAI published a paper on exploring data- Bavarian et al. (2022) Artificial Intelligence in the
2022 optimal models with FIM. Construction Industry
Sep 1st, OpenAI reduced the pricing of the GPT-3 Decoder (2022), F Architect 19 years Research and Development with
2022 model by 66% for davinci model. OpenAI (2022) expertise in business intelligence
Sep 21st, OpenAI announced Whisper (speech Radford et al. (2022) G AI/ML Engineer 10 years IT with expertise in developing
2022 recognition). solutions for AEC companies
28/Nov/ OpenAI expanded GPT-3.5 to text-davinci- Karhade (2022)
2022 003 with improved language generation
capabilities. of a variety of tasks such as language production, sentiment analysis,
Nov 30th, OpenAI announced ChatGPT. OpenAI (2022) text categorization, and question answering (Brown et al., 2020). Lan
2022
Mar 14th, OpenAI released GPT-4, the latest and highly OpenAI (2023)
guage generation is the process of creating coherent and fluent text,
2023 anticipated addition to the GPT series. whereas sentiment analysis is the examination of text sentiment, such as
whether it is positive or negative. Text categorization entails classifying
text into several groups, such as news articles or product reviews. GPTs
Table 2 are useful tools for natural language interaction because
Search query. question-answering creates replies to open-ended questions (Brown
et al., 2020). GPTs have revolutionized NLP by bringing remarkable
Search Category Search Query
fluency and coherence to writings produced by machines (Devlin et al.,
Construction “Construction industry” OR “architecture engineering and 2018; Radford et al., 2018). The potential bias available in the
Industry construction industry” OR “AEC industry” OR “AECO industry”
pre-training data, which could affect the precision and calibre of the
GPT “Generative Pre-trained Transformers” OR “Generative AI” OR
“GPT” OR “GPT-1” OR “GPT-2” OR “GPT-3” OR “InstructGPT” generated text, is one of the main challenges. Moreover, there may be
OR “ChatGPT” OR “Transformer*” OR “GPT-4” significant computational costs and challenges associated with the use
and training of GPTs (Mantel, 2023). Future opportunities for more
research will continue to fuel continuing efforts to improve the accuracy
grammar rules. The transformer network creates coherent and fluent and efficacy of these models, which will drive the dynamic character of
output while the attention mechanism enables the model to focus on the GPT sector. For instance, there might be a decrease in the compu
pertinent portions of the input text (Zhang et al., 2022). Transformer tational expenses related to creating and utilizing GPTs, which would
networks, feedforward neural networks, and attention processes make make them more affordable for smaller businesses (Hussin et al., 2023;
up the building blocks of GPTs (Hernández and Amigó, 2021). Massive Paaβ and Giesselbach, 2023).
volumes of text data are trained during the pre-training phase of GPTs,
allowing the model to learn broad language patterns that may be honed 3. Methodology
for particular tasks (Kotei and Thirunavukarasu, 2023). Pre-training
often involves unsupervised learning without labels or annotations. A qualitative research approach is adopted in this research to achieve
After pre-training, the model is adjusted for a variety of tasks to increase the aim of the study. This involved leveraging three sequential steps, as
the quality and accuracy of the text that is produced for that activity, depicted in Fig. 2, including a literature review, critical review, expert
such as language modelling, text categorization, or question-answering. discussion, and a case study.
GPT models may be fine-tuned to accomplish a range of tasks with
great accuracy (Ouyang et al., 2022; Wei et al., 2023). GPTs are capable
3
Fig. 2. Research approach.
i) Initial Exploration (Phase I): A search query (shown in Table 2) is Following the expert discussion, a critical review was conducted on
developed and used in conducting a detailed search in Scopus, ACM, the identified opportunities and limitations of GPT models in the AEC
Web of Science, Science Direct, and Google Scholar databases. These industry. A critical review entails a detailed analysis and critique of the
databases were selected because of their capacity, and relevance to work with the view of providing an objective assessment of the work,
the subject matter and have been well-adopted in similar studies implications, and insights. It is employed in this study to evaluate and
(Saka et al., 2023). The outputs from all the databases revealed that, identify the opportunities and limitations of GPT models in the AEC
although transformers have been gaining increased attention in the industry based on expert discussion and relevant extant studies.
AEC industry, the application of GPT models is low. As such, only
three publications were identified that specifically leveraged GPT iii) Use case (Phase III): One of the identified opportunities of GPT
models in the AEC industry. Furthermore, a subsequent search on models in Phase II – material selection and optimization – is
arXiv database suggested that there are 3 related preprints – manu evaluated in this study. The system architecture is proposed for
scripts that are yet to be formally peer-reviewed – available online. leveraging GPT for material selection and optimization which is
All the identified outputs were reviewed and citation tracking – subsequently verified and validated. Verification deals with
getting previous studies from references - was used till saturation. ‘building the product right’ and validation deals with ‘building
These outputs are tabulated in an Excel file and considered to avoid the right product’ (Boehm, 1984). The verification and validation
publication bias. Lastly, the outputs were reviewed, and themes process is done using checklists and case-testing approach (Saka
related to the objectives were identified and tabulated for usage in et al., 2022). The checklist technique involves the use of
the subsequent research phases. specialized lists based on experience to check significant issues
that are critical for product development whilst case-testing
ii) Expert Discussion and Critical Review (Phase II): Based on the out technique entails prototyping the developed products. As such,
puts of Phase I, expert discussion and a detailed critical review were the employed techniques in this study fulfil the basic criteria of
employed to complement the few research studies. Expert discussion verification and validation processes – completeness, consis
is a method of obtaining in-depth insights on a specific research tency, feasibility and testability (Boehm, 1984).
theme by facilitating a structured discussion with a group of panel
lists with diverse backgrounds and experiences. It is a strong 4. Findings
approach when the research area is new or ambiguous and there is a
need for the generation of innovative ideas, identifying problems, This section presents the findings from the literature search on the
develop recommendations and solutions to complex problems. It current applications of GPT models in the AEC industry and expert
could take various forms such as a Delphi survey or Focus Group discussion on opportunities and limitations.
Discussion and involves a moderator to ensure group dynamic and
round participation from all the experts (Hsu and Sandford, 2007;
Jenkins and Smith, 1994). As such, the employed expert discussion is 4.1. Current applications
a modified classical Delphi survey with panellists who are selected
based on predefined criteria - domain expertise in Artificial Intelli GPT models are still new in the construction industry, unlike other
gence and the AEC industry with a minimum of 10 years’ experience. LLMs such as BERT which has been gaining widespread applications in
Ten experts were identified and contacted for the research discussion the AEC industry since 2020. Only 5 papers that have applied GPT
and only 7 accepted the invites and participated in the research. This models were retrieved and summarized in Table 4.
is considered acceptable as a minimum of 7 is considered sufficient These extant studies have applied GPT models for question-
for the Delphi survey and this current study leverages the expert answering, information retrieval from BIM model and for scheduling
discussion to complement critical review and case study validation & sequencing tasks. There are inherent limitations as a result of the GPT
(Hon et al., 2011). Table 3 shows the demographic details of the models employed in the applications and limitations as a result of the
experts who participated in the discussion. The experts were con development approach employed in the studies. GPT models work best
tacted virtually, and an average discussion of 25 min was undertaken with well-structured and clean data, which is often unavailable in the
on opportunities and limitations of GPT models in the AEC industry. AEC industry; as such, there is a need to pre-process the data before
Thematic analysis and mapping were conducted to identify specific usage in GPT. Unstructured data are parsed into readable formats like
opportunities mentioned by the experts. Opportunities and limita Plain text files (.txt), Comma-separated values (.csv), and JavaScript
tions were tabulated and pass across back to all the panellists for final Object Notation (.json). For instance, Zheng et al. (2023) employed
review. BSON (Binary JSON) format by extracting building objects and prop
erties from BIM model which is subsequently cleaned and stored in
4
Table 4 Table 5
Current applications. Categorization of opportunities by experts.
S/ Application Purpose Access and GPT Limitations Phase Opportunities A B C D E F G
N model
Pre-Design Optimal design and ✓ ✓ ✓ ✓ ✓ ✓
(Specification)
construction
1 BIM-GPT ( Information API access and No techniques
Zheng and retrieval from BIM GPT-3.5-turbo quantitative Procurement ✓ ✓ ✓ ✓ ✓
Fischer, using natural (Temperature = evaluation Project brief and client ✓ ✓ ✓
2023) language 0) Single-turn requirements
conversation Lessons from project ✓ ✓ ✓ ✓ ✓
2 RoboGPT ( Automate ChatGPT-4 API Blackbox Project execution ✓ ✓ ✓
You et al., sequence planning access (Not High risk planning
2023) of construction provided) Unable to Project management ✓ ✓ ✓
tasks in robot- leverage visual and planning
based assembly information Design Generation of design ✓ ✓
3 Prieto et al. Scheduling of ChatGPT-3.5 (Not Zero-shot concept
(2023) construction tasks provided) learning Regulatory compliance ✓ ✓
No Material selection and ✓ ✓ ✓ ✓ ✓
quantitative optimization
evaluation Quantity take-off and ✓ ✓ ✓ ✓ ✓
4 (Uddin et al., To support hazard ChatGPT-3.5 (Not Zero-shot costing
2023) recognition and provided) learning Improving energy ✓ ✓ ✓ ✓
construction safety No efficiency analysis
education quantitative Design specification ✓ ✓
evaluation Construction Scheduling and ✓ ✓ ✓ ✓ ✓ ✓ ✓
5 Amer et al. Integrating master GPT-2 (small Large data set logistics
(2021) schedules with version) Regulatory compliance ✓ ✓ ✓ ✓ ✓ ✓ ✓
look-ahead plans Risk identification, ✓ ✓ ✓ ✓ ✓ ✓ ✓
assessment, and
management
MongoDB. On the other hand, Uddin et al. (2023), Amer et al. (2021), Progress monitoring ✓ ✓ ✓ ✓
and report
Prieto et al. (2023) and You et al. (2023) leveraged text for the appli
Site safety ✓ ✓ ✓ ✓ ✓ ✓ ✓
cation. Despite some of the limitations, the studies are important and management
contribute to the new area of GPT applications in the AEC industry and Resource allocation ✓ ✓ ✓ ✓ ✓ ✓ ✓
provide a basis for new studies to build on. and optimization
Change order ✓ ✓ ✓
management
4.2. Expert discussion Quality control and ✓ ✓
assurance
Table 5 summarizes the opportunities for GPT models in the AEC Documentation ✓ ✓ ✓ ✓ ✓ ✓ ✓
industry as identified through expert discussion, and they are catego Dispute resolution ✓
Budgeting and cost ✓
rized into different phases of the project lifecycle. planning
Table 6 presents the challenges and limitations of deploying GPT Operation and Predictive ✓ ✓
models in the AEC industry. Some of these limitations are inherent Maintenance maintenance
limitations of GPT models, whilst others are industry-specific challenges Energy management ✓ ✓
and optimization
facing the deployment of GPT models in the AEC industry.
Incident and resolution ✓ ✓ ✓ ✓ ✓ ✓ ✓
The following Sections 5 and 6 provide a detailed discussion of the Lifecycle management ✓ ✓ ✓ ✓ ✓ ✓ ✓
opportunities and limitations identified in the previous section. Op of asset
portunities and limitations of similar themes are synthesised and dis Occupant ✓ ✓ ✓ ✓ ✓ ✓ ✓
cussed. These opportunities require different levels of development and communication and
support
interaction with the GPT models ranging from zero-shot learning to Regulatory compliance ✓ ✓ ✓ ✓ ✓ ✓ ✓
finetuning and integration with different knowledge sources. Similarly, management
some of the limitations are inherent in the GPT models whilst others can Space/facility ✓ ✓ ✓ ✓ ✓ ✓ ✓
be attributed to the construction industry context. management
Performance ✓ ✓
monitoring
5. Opportunities Sustainability ✓ ✓
Waste management ✓ ✓
The opportunities are categorized into different phases of the con and recycling
struction project lifecycle – predesign, design, construction, operation & Demolition Demolition protocol ✓ ✓ ✓ ✓ ✓ ✓ ✓
Waste management
maintenance, demolition phase and value-added services. These op Redevelopment plan ✓ ✓ ✓ ✓ ✓ ✓ ✓
portunities are made possible by the inherent capabilities of GPT Regulatory compliance ✓ ✓ ✓ ✓ ✓ ✓ ✓
models, integration of GPT with existing systems and leveraging and permit
plugins/add-ins that extend the functionalities of GPT. The following Costing ✓
Environmental impact
subsections present these opportunities and how GPT models can be
✓ ✓ ✓
analysis
leveraged. Material recovery ✓ ✓
Risk assessment ✓ ✓
5.1. Pre-design phase Value added Knowledge ✓ ✓ ✓ ✓ ✓ ✓ ✓
management and
training
Pre-design construction activities serve as the foundation for project Customer services ✓ ✓ ✓ ✓ ✓ ✓ ✓
success, budgeting, and optimization. They comprise site study, pro (continued on next page)
gramming, construction cost analysis, and value engineering. As seen in
5
Table 5 (continued ) Also, effective prompts can be written using ‘Act As’ approach to engage
Phase Opportunities A B C D E F G the GPT model in providing technical information and clarification on
optimal design and construction techniques. For instance, the ‘Act As’
Stakeholder ✓ ✓ ✓
communication
was employed as shown in Fig. 4.
Business intelligence ✓ ✓ ✓ ✓ ✓ ✓
Conversational AI/ ✓ ✓ ✓ ✓ ✓ 5.1.2. Procurement decision support
Chatbot In the predesign phase, current procurement decision support pro
cedures include techniques such as value engineering, life cycle costing,
and multi-criteria decision-making (MCDM) (Tezel and Koskela, 2023).
Table 6 The efficacy of these strategies, nevertheless, is restricted as they rely on
Limitations of GPT in the construction industry. specialized knowledge, which is generally subjective and susceptible to
Limitations A B C D E F G human fallibility (Ratnasabapathy and Rameezdeen, 2010). Further
more, previous studies have underscored the necessity for enhanced
Hallucination ✓ ✓ ✓ ✓
Data ✓ ✓ ✓ ✓ ✓ data processing and analysis to optimize decision-making, particularly
Expertise knowledge ✓ ✓ ✓ ✓ ✓ ✓ considering the escalating quantity of project data acquired during the
Intellectual property and confidentiality ✓ ✓ preliminary design stage (Budayan et al., 2015). As posited by McBride
Safety ✓ ✓ ✓ ✓ ✓ ✓ ✓
et al. (2021), GPT models present a fresh approach to augmenting pro
Trust and acceptance ✓ ✓ ✓ ✓ ✓ ✓ ✓
Accountability and liability ✓
curement decisions in the preliminary design phase by surmounting the
Ethics ✓ ✓ ✓ ✓ ✓ ✓ ✓ limitations associated with conventional methodologies. GPT models,
Skills and training ✓ ✓ ✓ ✓ ✓ with their powerful natural language processing and machine learning
Interoperability and integration ✓ capabilities, can evaluate data, delivering insightful suggestions that
Capital/cost
encourage accurate, data-driven decisions (OpenAI, 2023a,b). GPT
✓ ✓ ✓ ✓ ✓ ✓ ✓
Infrastructure requirements ✓ ✓ ✓
Scalability ✓ ✓ ✓ models also promote more efficient cooperation among stakeholders
Performance optimization ✓ ✓ ✓ through real-time information and predictive analysis (Abioye et al.,
Cybersecurity ✓ ✓ ✓ ✓ 2021; Momade et al., 2021). Prompts can be developed allowing pro
Interdisciplinary ✓ curement decision-makers to swiftly acquire important information,
Cultural and social consideration ✓ ✓ ✓ ✓
Latency issue
explore options, and make educated decisions. GPT models can be used
in performing textual analysis to provide insights based on supplier re
views, and product description historical purchase data for informed
Fig. 3, this section examines the status of pre-design in building projects decisions. Also, an integrated system can be developed with BIM, GPT
and the potential benefits of utilizing GPT models. model and procurement database. This integrated system can be lever
aged for queries related to procurement and automating workflow such
5.1.1. Expert guidance on optimal techniques for design and construction as order placements and status updates.
Stakeholders collaborate throughout the pre-design stage of con
struction to establish project objectives. Pre-design’s purpose is to 5.1.3. Development of project brief and client requirements
identify critical aspects and restrictions that will guide subsequent A crucial part of the predesign stage of the construction process is the
design and development processes. Currently, the predesign landscape is development of the project brief and client requirements, often known
defined by a mix of traditional approaches and specialized technologies, as Employer Information Requirements (EIR) in BIM projects (Catenda,
each with its own set of benefits and drawbacks. 2020; Kim et al., 2022). Creating a thorough project brief and outlining
Due to cognitive biases and information overload, the traditional the client’s requirements in detail are essential stages that provide the
reliance on expert knowledge and experience might result in less-than- groundwork for the whole construction process at the pre-design stage of
ideal outcomes (Levy, 2010). Extant studies have documented the the project. Multiple methods and constraints, such as the use of tem
importance of predesign in reducing project risk, optimizing resource plates, checklists, and questionnaires, are part of the existing ways of
allocation, and enhancing overall project success (Guo and Zhang, 2022; collecting design requirements (Assaf et al., 2023).
Lu et al., 2020). This phase can further be enhanced and transformed by Numerous approaches and restrictions that affect the creation of the
GPT models, which provide expert guidance on the best design and project brief and employer information requirements (EIR) have been
building procedures, resulting in higher productivity, decreased costs, highlighted by recent studies. By automating some procedures and
and improved sustainability (OpenAI, 2023a,b). Users may access the fostering cooperation and communication between project teams and
model using the user interface or API access to query its vast knowledge clients, the emergence of GPT models has the potential to completely
base on best practices in design and construction (Smith et al., n.d.). transform the development of project briefs and EIRs (Shaalan, 2010).
Fig. 3. Opportunities for GPT models in the Pre-design Phase.
6
Fig. 4. Act-As prompting.
For instance, GPT model can be integrated with an existing system for definition, scheduling, resource allocation & estimation, risk manage
interacting with and supporting clients in the development of initial ment and decision support system. For instance, with prompting, GPT
project briefs (Saka et al., 2023). Fig. 5 shows the questions that the GPT models can help understand the impact of diverse factors on projects and
model would ask the client in a conversational style to automatically provide recommendations for effective management. Also, GPT models
develop the initial project brief. Based on the develop brief, the GPT have been proved applicable in the identification of tasks and scheduling
model can provide design suggestions and technology recommenda based on specific project requirements (Prieto et al., 2023; You et al.,
tions. Also, the brief can be compared with regulatory compliances for 2023). As such, GPT models could also be leveraged for analysing data
non-compliance and potential risks can be highlighted in the brief. and providing relevant insights to the project manager during the pre
design stage. Similarly, with zero-shot learning, GPT can act as a re
5.1.4. Project management planning pository for industry standards and lessons learnt from previous projects
Project management planning is vital to the successful delivery of for stakeholders to make informed decisions.
projects, as poor planning could lead to cost & time overrun and quality
issues (Asiedu et al., 2017). It entails inputs from project teams and
5.2. Design phase
stakeholders and sets out details of how the project is to be executed,
monitored, and controlled. As such project management planning
This section presents the opportunities for leveraging GPT models
components are baselines for scope, schedule, cost, requirement man
during the design phase of construction projects as shown in Fig. 6.
agement plan, change management plan, configuration management
plan and process improvement plan (Globerson and Zwikael, 2002).
5.2.1. Generation of design concept
GPT models trained on corpus of database could be finetuned to
An essential step in the design process is creating design options that
support project management planning during predesign stage in scope
satisfy project requirements. In the past, architects and engineers have
Fig. 5. Conversational development of Client’s brief.
7
Fig. 6. Opportunities for GPT models in the Design Phase.
relied heavily on their personal knowledge and previous experience to making factors for material selection (Florez and Castro-Lacouture,
manually develop design concepts, which can be lengthy and subjective 2013). During the selection, the designer would need to consider all
(As et al., 2018; Lewis and Séquin, 1998). Recently, GPT have shown these factors coupled with other subjective and objective measures to
promise for accelerating and enhancing this process through their ability reach a decision which may not be the optimal solution. As such, extant
to generate coherent, context-relevant text and images after training on studies have proposed the use of different optimization approaches such
large datasets of domain-specific information. For example, by learning as mixed integer optimization, fuzzy logic approach, and Grey relational
from architectural plans, construction codes, and design guidelines, GPT analysis to solve the problem (Emovon and Oghenenyerovwho, 2020).
models can acquire construction industry language patterns and GPT models can be leveraged for optimization of material selection with
leverage that knowledge to automatically generate design ideas tailored detailed consideration of different factors and parameters. This can be
to given constraints (Zheng and Fischer, 2023). Lastly, with the inclu integrated with BIM model to optimize material selection during design
sion of DALL-E, GPT model can be used in generating design ideas with and to provide design alternatives. It would involve evaluating the
zero-shot learning as shown in Fig. 7. material of building components (from BIM) based on material prop
erties and performance database (from GPT) to fulfil predefined criteria.
5.2.2. Automated regulatory compliance This is further demonstrated in Section 7 for validation.
Complying with regulatory requirements is a critical aspect of the
design phase, as failure to meet these standards can result in legal issues, 5.2.4. Quantity take-off and costing
delays, and safety hazards. The complexity and frequent updates of Quantity take-off and costing form an integral part of successful
building codes and regulations make it challenging for design pro project delivery, and the quantity surveyor is often saddled with this
fessionals to ensure compliance manually (Dimyadi et al., 2015). responsibility (Saka and Chan, 2019). GPT models can be leveraged for
Nevertheless, GPT models can play a significant role in automating quantity take-off and costing of building projects by providing detailed
regulatory compliance checks, reducing errors, and streamlining the information about the project – design, material and other specifica
design process. This can be achieved by analysing architectural and tions. This can be achieved by providing textual data for the model with
structural designs and comparing them against relevant building codes. necessary cost databases and estimation methods. Based on these,
Conversely, the model can identify potential issues, allowing designers prompts can be developed for quantity take-off of the project with
to rectify them before construction, thereby saving time and minimizing elemental breakdown and subsequent costing of the quantities. Also,
costly revisions. GPT can be used to prepare bills of quantities and analyse past bills of
quantities to draw insights and make predictions. Machine learning
5.2.3. Optimizing material selection models can be developed in GPT model using past cost data which can
The construction industry consumes a lot of resources and contrib then be used in predicting the cost of new projects. Lastly, GPT can be
utes significantly to the green gas emissions. Material selection is one of queried to provide steps to be followed using python in extracting
the ways to reduce the environmental impacts of a project during its quantities from IFC file for quantity take off.
lifecycle. Factors such as cost constraints, location of the component,
design consideration and environmental requirements are decision- 5.2.5. Improving energy efficiency analysis
There has been a surge in energy demand with rapid development
over the decades in the construction industry (Oluleye et al., 2022). The
energy efficiency of buildings are is broadly based on the building en
velope, which influences energy consumption and the rate at which
energy is lost in the building (Abu Bakar et al., 2015). GPT models can be
provided with information on standards, regulations, passive design
principles, building facades optimization and renewable energy systems
for it to be leveraged for improving energy efficiency analysis. As such,
GPT models can provide guidance on selection of simulation tools,
interpretation of results, identify opportunities for improvement (such
as optimizing building orientation, insulation, energy systems) and
analysing cost-effectiveness of proposed solutions. Also, life cycle
Fig. 7. Image Generation from ChatGPT using DALL-E.
8
analysis can be computed with prompts developed for usage on the integer programming (MILP) and machine learning (ML) to address
finetuned model and documentations can also be prepared by the GPT these difficulties, particularly emphasizing the utilization of long short-
models. Whilst some of these are possible with zero-shot learning, to get term memory (LSTM) models (Al-shihabi & Mladenović, 2022). How
best output from the GPT models, there is need to fine-tune the model ever, GPT models exhibit certain advantages when compared to con
and integrate it with other BIM and energy efficiency analysis tools. ventional techniques. Unlike MILP and LSTM models, which primarily
operate on structured and numerical data, GPT models possess the
5.2.6. Design specification ability to effectively acquire and analyse textual information, thereby
As one of the largest industries with in-built complexities, design facilitating a more comprehensive understanding of the complexities
specification plays a crucial role in the success of any project within the associated with construction project scheduling and logistics. In contrast
construction industry. Design specification, being a technical document, to the mathematical optimization problems described in MILP models,
provides all necessary information for the design phase of a construction GPT models leverage learned patterns and representations derived from
process. The existing process for developing a design specification the available data. This approach reduces the need for explicit mathe
document entails manual work, which can be time-consuming and matical modeling and enables a more flexible and intuitive approach to
laborious. However, with the introduction of the GPT models, new addressing scheduling and logistical challenges in construction en
prospects for automating the process of developing design specification deavors. Furthermore, the utilization of GPT can be employed to
documents have surfaced. implement the Critical Path Method (CPM) and Program Evaluation and
Since GPT models can generate logical, grammatically correct lan Review Technique (PERT) by inputting the project schedule.
guage (Brown et al., 2020), they have the potential to automatically
construct complete design specification documents when provided with 5.3.2. Automated regulatory compliance implementation
key inputs. Engineers and architects would simply need to feed the The intricacy of regulatory compliance is a significant barrier within
model the core information needed, and it could produce the full spec the building business. It is of utmost importance for construction en
ifications (Hayman, 2022; Parm AG, 2023). This has the potential to terprises to ensure adherence to a multitude of legislative requirements
greatly reduce the effort and resources required to develop specifica and performance-based rules (Zhang and El-Gohary, 2016). The con
tions. Additionally, GPT-generated specifications may have higher ac ventional approaches to regulatory compliance verification are charac
curacy and consistency than manual approaches. terized by their tendency to consume significant amounts of time, their
susceptibility to errors, and their heavy reliance on resources. The
emergence of artificial intelligence (AI), particularly GPT models, pre
5.3. Construction phase sents potential avenues for enhancing efficiency and automating the
regulatory compliance procedure. GPT models possess the potential to
This section presents the opportunities of leveraging GPT models streamline regulatory compliance in the construction sector by virtue of
during the construction phase as shown in Fig. 8. their natural language processing (NLP) capabilities and aptitude for
comprehension. Beach et al. (2020) argue that these models establish
5.3.1. Managing project schedule and logistics the capacity to assess and isolate relevant information from textual in
The process of scheduling entails the development of a comprehen formation pertaining to construction regulations. This process involves
sive plan that outlines the sequence of activities, tasks, necessary re changing the extracted information into logic clauses, which can then be
sources, and estimated time required for completion (Castro-lacouture utilized for automated reasoning. The implementation of automation in
et al., 2009). On the other hand, logistics management encompasses the the compliance checking process can yield substantial reductions in time
deliberate formulation, implementation, and regulation of the trans and effort needed for regulatory compliance evaluation. GPT models, in
portation of individuals, resources, and machinery to and from the particular, have the potential to function as valuable aids in the realm of
construction site (Dannoun, 2022). The successful execution of a con automated regulatory compliance.
struction project hinges upon the efficient handling of scheduling and
logistics. Managing logistics and scheduling are difficult responsibilities 5.3.3. Risk identification, assessment, and mitigation
since the supply chain is unpredictable and complex. The proactive identification and resolution of risks can effectively
Previous research has employed a hybrid approach to connect mixed
Fig. 8. Opportunities for GPT models in Construction Phase.
9
mitigate any negative impacts on project objectives. The potential vital during the construction phase to prevent accidents, reduce injuries,
hazards linked to a specific undertaking may involve a range of factors, and ensure compliance with safety regulations. However, traditional
including external factors, internal intricacies, technical challenges, or approaches to site safety management can be time-consuming and
unforeseen situations. During the construction period, there exist labour-intensive (Tixier et al., 2016). The integration of GPT in site
various possible sources of dangers that may manifest. The factors safety management presents new opportunities for improving safety
encompassed in this context are project complexity, resource availabil practices during the construction phase. By leveraging the capabilities of
ity, safety threats, design modifications, and contractual considerations GPT, construction companies can efficiently perform safety audits,
(Siraj et al., 2019). Risk management strategies play a crucial role in the automated risk assessments, and get insights into potential hazards
construction sector, however they are not without inherent limitations. (Porter, 2021).
Given these limitations, it becomes apparent that GPTs hold signif GPT can also assist in identifying high-risk areas on construction sites
icant potential in augmenting the identification, assessment, and man by analyzing text-based data like project documents, safety reports, and
agement of hazards within the construction industry (Mills, 2001; worker feedback to identify potential areas of concern that may not be
OpenAI, 2023a,b). GPT models provide the capacity to be employed for readily apparent. Also, image-based data can be evaluated by GPT to
comprehensive and flexible risk identification and assessment by scru identify hazards on construction sites as shown in Fig. 9. As a result,
tinizing substantial project documents and historical data, while also propose mitigation measures based on these historical data, documented
being capable of integrating new information as the project progresses. industry best practices, and regulatory guidelines (Ezelogs, 2023; Togal.
The aforementioned skill enhances the current comprehension of the AI, 2023). Moreover, GPT can play a crucial role in knowledge sharing
project’s risk assessment (Miller, 2016; OpenAI, 2023a; Zheng and and training. Construction workers can access GPT-powered platforms
Fischer, 2023). Predictive analytics holds the capacity to anticipate and for interactive safety training, where they can receive personalized in
evaluate probable hazards and their associated consequences, hence structions and guidance on safe practices (Uddin et al., 2023). GPT’s
augmenting the overall accuracy of risk profiles (Cornwell et al., 2022). NLP capabilities enable effective communication and prompt responses
In summary, it can be inferred that GPTs possess the capacity to auto to worker queries, enhancing the overall safety culture on construction
mate the generation of risk reports and various types of communication. sites. Additionally, GPT can facilitate incident reporting and analysis by
The implementation of this approach holds promise for improving the automatically categorizing incidents, identifying root causes, and rec
accuracy and consistency of risk communication, hence mitigating the ommending preventive measures, thereby enabling proactive safety
risk of misunderstandings and exclusions. management.
5.3.4. Project progress monitoring and reporting 5.3.6. Resource allocation and optimization
The act of monitoring a project during the construction phase is vital Resources on construction projects include plant, personnel, and
in order to ensure its effective culmination. The utilization of effective items necessary to complete tasks. Resources allocation is important and
monitoring strategies facilitates timely interventions and adjustments, can affect the delivery of projects in terms of cost, time and quality. The
ultimately leading to improved results. Project reports play a crucial role allocation of resources is dependent on the diverse features – nature of
in effectively conveying the present state, accomplishments, obstacles, the project and other key attributes (Kusimo et al., 2019). Approaches
and other significant facets of a project to stakeholders who possess a such as genetic algorithm and particle swarm optimization have been
vested interest in its ultimate results (Ibrahim et al., 2009). However, applied in optimizing resource allocation and levelling. GPT models can
conventional methodologies for monitoring and reporting project also be used in resource allocation and optimization by providing
progress possess inherent constraints. These approaches often rely on detailed information about the project. The models can provide rec
human judgment, introducing subjectivity and potential mistakes into ommendations by taking cognisant of the project requirements, avail
the process. Furthermore, it is important to acknowledge that these able resources, project duration, and critical path potential challenges.
methodologies possess the capacity to need a substantial investment of Also, scenario analysis can be conducted with the GPT model and
time and may not provide instantaneous or current insights into the documentation of the resources’ allocation.
progress of the project. The absence of immediate access to information
can hinder the ability to make proactive decisions, as emphasized by 5.3.7. Change order and quality management
El-Omari and Moselhi (2009). GPT models can assist the project team when there are changes in the
Prior studies have recognized that applying computer vision (CV) project scope by providing aspects of the project such as cost, time and
and machine learning (ML) models could potentially solve the problems quality that could be impacted (impact analysis) (Abioye et al., 2021).
related to progress reporting. The theoretical basis in question is sub The models can be employed in reviewing change orders to ensure they
stantiated by the empirical inquiries carried out by Ibrahim et al. (2009) are consistent with the contract documentation and for highlighting
and Kim et al. (2013). The utilization of sophisticated algorithms and discrepancies in the change order. Also, the models can be leveraged in
image processing methodologies is a characteristic feature of these the documentation of change order requests, communication and
models, enabling the extraction of significant insights from construction providing recommendations in the event of negotiation and dispute
data encompassing images, videos, and sensor data. GPT models exhibit resolution.
a notable advantage in this specific field owing to their employment of Quality control and assurance ensure that the quality of a project is in
robust language models that incorporate CV, ML, and NLP techniques. tandem with objectives and established standards whilst also mitigating
GPT models have demonstrated their capacity to comprehend, interpret, errors and defects. GPT models can be employed in planning and
and proficiently generate text that closely resembles human language. organizing inspection and testing activities based on defined project
Hence, it can be posited that these models possess the capacity to requirements and familiarization with principles (quality control and
augment the effectiveness of progress monitoring and reporting through assurance in the construction industry). GPT can be used to identify
the examination of textual data pertaining to projects, encompassing anomalies in quality data, suggest best practices standards for preven
project updates, documents, and reports. Therefore, GPT models can tive and corrective actions and assist in quality documentation (man
offer a complete solution for automating reporting and progress moni agement, technical and general procedure manual and policy manual).
toring during the building stage.
5.3.8. Construction project documentation
5.3.5. Site safety management Managing project documentation requires significant effort to ensure
The construction industry is associated with high rates of accidents that all the relevant documents are up-to-date, accurate, and accessible
and fatalities (Dolphin et al., 2021). Managing site safety is, therefore, to all relevant stakeholders (El-Omari and Moselhi, 2009). This is often
10
Fig. 9. Hazard identification using computer vision capability of ChatGPT.
done traditionally, which is largely dependent on human judgments, 5.4.1. Predictive maintenance
thereby resulting in the infrequent provision of accurate and timely A considerable challenge in the operational stage of a construction
construction data (Ibrahim et al., 2009). Hence, the experts have high project lies in efficiently overseeing maintenance activities. Conven
lighted the adoption of GPT to mitigate this issue, as the models can be tional practices in this area have been more reactive than proactive,
trained to recognize and analyse different types of documents or re acting only after an equipment breakdown. Such an approach frequently
cordings, extract relevant information from them, and create a central results in unanticipated periods of inactivity, increasing costs for repairs,
ized database for easy access by all stakeholders. As such, progress and the suboptimal distribution of resources (Bouabdallaoui et al.,
reports or status updates can be easily generated based on the extracted 2021).
data from the project documentation. One of the ways that GPT models can transform maintenance prac
tices is by enabling predictive maintenance. Predictive maintenance is a
5.3.9. Claim and dispute resolution data-driven approach that can foresee equipment failures or perfor
The financial health and integrity of construction stakeholders can mance problems before they happen (Taiwo et al., 2023). GPT models
be affected by the occurrence of claims and dispute resolution; hence, it can analyse historical data, sensor readings, and other relevant factors to
is a critical aspect in the construction industry that needs optimum find patterns and indicators of potential failures, thus allowing main
attention. Disputes often arise due to various reasons, such as design tenance teams to take preventive actions, plan repairs or replacements at
errors, construction delays, or cost overruns, and can result in lengthy the best times, and reduce the impact on operations. It can help to lower
and costly legal battles (Dikbas et al., 2010). maintenance costs by preventing unnecessary repairs and prolonging
Drawing upon past records, such as modifications to orders and the equipment’s lifespan.
schedules for projects, GPT models possess the capability to forecast the
odds of encountering conflicts or claims in current or forthcoming pro 5.4.2. Energy management and optimization
jects. Additionally, these models can propose varied strategies for Buildings and infrastructure heavily consume energy, and inefficient
settling disputes tailored to the unique conditions surrounding the usage increases costs and environmental impact. Traditional manual
conflict. Achieving this is possible by educating the models on an array energy monitoring methods are labor-intensive, resource-demanding,
of methods for resolving disagreements, encompassing techniques like and error-prone (Hagras et al., 2008). However, GPTs can optimize
mediation, arbitration, and negotiation, thus equipping them to examine energy management by analyzing large datasets - like historical usage,
the particulars of a conflict and recommend suitable avenues for its weather, occupancy, and equipment data - to identify savings opportu
resolution (Chaphalkar et al., 2015). nities (Zheng and Fischer, 2023). GPT-enabled energy management
provides several benefits: real-time tracking to detect anomalies and
5.3.10. Project budgeting and cost planning waste quickly, predictive demand forecasting for resource planning, and
Efficient budgeting and cost planning are vital during a construction customized recommendations for energy saving measures like adjusting
project’s lifespan, as the industry is known for frequent overspending heating, ventilating, and air conditioning (HVAC) settings, or equipment
and budget deviations. These budget misrepresentations substantially upgrades tailored to the building.
impact finances and timelines (Wang et al., 2012). Traditional budget
ing relies on expert judgment, which tends to be subjective and 5.4.3. Incident reporting and resolution
error-prone (Cheng et al., 2009). In contrast, GPT models have the In the operational and maintenance phases of construction activities,
capability to automatically generate accurate and data-driven financial efficient procedures for reporting and resolving incidents are critical for
forecasts and estimates based on historical records and industry norms. the safeguarding of safety, asset functionality, and overall operational
It should be acknowledged that the precision of the model depends on performance. Events such as equipment failures, safety risks, and in
the prompt quality and the data employed. terruptions in operation are not uncommon and necessitate immediate
and judicious action to limit their consequences. Traditional methods for
5.4. Operation and maintenance phase incident reporting and resolution often involve manual systems, paper
documentation, and elongated communication channels, contributing to
This section presents opportunities for leveraging GPT models in the delayed responses and potential communication errors (Bach et al.,
operation and maintenance phase as shown in Fig. 10. 2013).
11
Fig. 10. Opportunities for GPT models in Operation and Maintenance Phase.
GPT models represent a promising approach for addressing these Chatbot to collect and sort occupant maintenance requests with zero-
challenges. Exploring its NLP and ML functionalities, these models have shot learning.
the capability to efficiently evaluate incident reports, recognize recur This can be deployed on the FM company’s website or an automated
ring themes, and classify incidents based on their level of criticality and voice assistant. For the automated voice assistant, Automatic speech
time sensitivity. Moreover, GPT models can facilitate the creation of recognition (ASR) and Text to speech synthesiser will be required. ASR
standardized incident reports, thereby capturing essential details and will convert speech to text for interaction with the FacilityBot and TTS
enabling uniform documentation. will convert the text from the system to voice (Saka et al., 2023).
5.4.4. Lifecycle management of asset 5.4.6. Regulatory compliance management

Conventional asset management approaches tend to be intricate, GPT models can be trained with relevant compliance documents,
time-consuming, and financially burdensome (Grussing, 2014). In this guidelines, and reporting requirements to capture patterns and
regard, GPT models offer a compelling solution by streamlining asset embedded knowledge. These models can be used in compliance evalu
management through automated processes encompassing data collec ation of facility management activities by comparing new inputs with
tion, analysis, and decision-making. By harnessing the power of GPT regulatory requirements to identify non-compliance issues. Similarly,
models, it becomes possible to accurately predict an asset’s remaining the GPT models can be employed in documentation by highlighting non-
useful life, empowering maintenance teams to proactively schedule compliance issues, providing recommendations, and creating stan
necessary maintenance and repairs, thereby mitigating the risk of un dardized reports to save time and ensure consistency as shown in Fig. 12.
expected failures. The proactive approach facilitated by GPT models
significantly reduces the need for costly emergency repairs and mini 5.4.7. Space optimization and performance monitoring
mizes downtime, resulting in substantial cost savings. GPT models can be leveraged to analyse information about spaces
available in a facility to allocate the space for the best usage based on
5.4.5. Occupant communication and support predefined objectives to maximize productivity and utilization. Simi
GPT models can be employed as a conversational system between the larly, data from sensors can provide information about real-time usage
occupant and the facility management (FM) department. These can be to identify peak periods, usage patterns, and congestion areas to opti
prompted to act as chatbots for communicating and supporting facility mize space layout, and streamline workflow. This would also enable FM
users in real-time whilst allowing the facility management personnel to personnel to make space reconfiguration decisions, plan for accommo
attend to meet their requests faster and more effectively. For instance, dating future growth, and adapt to seasonal changes and new work
GPT models can collect maintenance requests, general enquiries from trends.
facility users to ensure clear and consistent communication and neces Key performance Indicators (KPIs) such as equipment uptime, user
sary support. Fig. 11 shows ChatGPT prompted to serve as a Facility satisfaction, space utilization, energy efficiency, maintenance response
12
Fig. 11. Facility chatbot.
Fig. 12. Compliance report generation.
time and energy efficiency are some of the metrics employed to measure construction industry, and stakeholders are increasingly focusing on
performance during the facility management phase of construction measuring, reporting, and improving the environmental, social, and
projects. Information such as maintenance records, equipment perfor economic performance of buildings and infrastructure (Marjaba and
mance, energy consumption can be obtained from sensors, project Chidiac, 2016; Taiwo et al., 2023a). Nevertheless, thoroughly doc
management and energy management systems for usage of GPT models umenting sustainability practices and metrics is challenging due to
in performance monitoring. Also, real-time data would enable the FM extensive, complex data collection and analysis requirements.
personnel to track performance and deviations from desired perfor GPT models could streamline sustainability reporting by extracting
mance levels. The developed KPIs can be compared to similar projects insights from large datasets - like energy use, occupant feedback,
and best practices can be recommended by the GPT models. Similarly, maintenance records and environmental monitoring data. When trained
performance forecast can be conducted based on historical performance on sustainability rating systems like LEED or BREEAM (Yeung et al.,
data and GPT models can be leveraged for documentation of facility 2020), GPTs would help generate automated reports, identifying per
performance in accordance with defined standard to ensure consistency formance gaps and recommending concrete improvements to facilities’
and saves time. sustainability.
5.4.8. Sustainability reporting and improvement 5.4.9. Waste management and recycling
Sustainability has become a paramount consideration in the Effective waste management and recycling play a pivotal role in the
13
operation and maintenance phase of the construction industry. Properly intelligent decision support to emergency responders. Leveraging their
managing construction and demolition waste, as well as responsibly analytical capabilities, these models can assess the severity of the situ
handling ongoing operational waste, is vital for minimizing environ ation, identify potential hazards, and recommend appropriate actions
mental impact and fostering sustainable practices (Adedara et al., 2023). such as evacuation, containment, or fire suppression. Their insights and
Nevertheless, waste management poses a challenge, demanding guidance assist responders in making informed decisions and executing
streamlined processes for efficient tracking, sorting, and disposal effective emergency plans. The post-incident management can also be
methods (Amaral et al., 2020). enhanced by evaluating incident reports, witness statements, and other
Due to their intelligent decision support and optimization capabil relevant documents to identify root causes, lessons learned, and op
ities, the experts have identified GPT models to offer valuable contri portunities for improvement (Beata et al., 2018).
butions to improving waste management and recycling practices. By
training these models on waste management regulations, best practices,
and recycling guidelines, they can effectively analyse waste-related data 5.5. Demolition phase
and provide informed recommendations. An example of their utility is
waste sorting and classification. GPT models can be adapted to analyse This section elucidates the potentialities for harnessing GPT models
images or descriptions of waste materials and based on their training on during the construction demolition phase, as visually depicted in Fig. 13.
extensive datasets of waste items and their appropriate recycling or
disposal methods, the models can accurately identify the correct 5.5.1. Development of demolition protocol
handling procedures for various waste types. This capability empowers The demolition phase of a building endeavour involves the careful
facility managers to streamline waste management processes and and controlled removal of existing structures in a manner that ensures
enhance recycling rates, ultimately contributing to more efficient and safety and security. Chew (2010) asserts that the achievement of project
sustainable practices. success is contingent upon the scrupulous formulation of plans and the
flawless implementation thereof, while duly considering the welfare of
5.4.10. Emergency response during fire or other hazards labourers, the preservation of the environment, and the overarching
During fire outbreaks or other hazardous incidents, the ability to objectives of the project. The establishment of a comprehensive demo
respond swiftly and effectively is vital for protecting occupants, mini lition procedure is of utmost importance to provide guidance for de
mizing property damage, and restoring normal operations promptly. molition operations and effectively address potential hazards.
Emergency situations are often complex and dynamic, demanding rapid GPT models possess the potential to contribute significantly to the
decision-making and coordinated efforts among stakeholders (Jiang, advancement of demolition protocols through their proficiency in data
2019). processing and risk assessment. The models possess the capability to
These complex situations can be addressed using GPT models. This is effectively handle substantial amounts of data, encompassing project
achievable by training the models on diverse datasets, including infor plans, site circumstances, and historical demolition records. Their pri
mation on past fire incidents, emergency protocols, and safety regula mary function is to detect potential dangers, establish the most advan
tions. Subsequently, they can analyse real-time data and provide tageous demolition sequences, and propose suitable safety measures.
One possible application of GPT models is to predict and evaluate risks
Fig. 13. Opportunities for GPT models in the Demolition Phase.
14
in various domains. Furthermore, via the incorporation of variables like construction teams. Furthermore, it is crucial to recognize that the
as site accessibility, waste management techniques, and equipment explication of regulations and requirements may give rise to variations,
utilization, these models have the capacity to develop demolition plans thereby resulting in inconsistencies and potential concerns regarding
that effectively limit operational interruptions, mitigate environmental compliance.
consequences, and optimize the allocation of resources. GPT models exhibit an inherent capacity to perform thorough anal
ysis and interpretation of regulatory documents, guidelines, and codes,
5.5.2. Waste recycling and safe hazardous material disposal thereby facilitating the timely provision of guidance and ensuring
Demolition projects yield a substantial quantity of refuse, including rigorous compliance with relevant regulations. These individuals exhibit
fragments, building components, and conceivably perilous compounds. the capability to facilitate the identification of the specific permits and
Effectively handling of waste, in conjunction with the adoption of suit approvals required for each demolition project, thereby expediting the
able recycling and disposal methods, assumes a pivotal role in advancing permit application process. The utilization of GPT models in the exam
environmental sustainability and complying with regulatory obligations ination of regulatory prerequisites presents a promising opportunity to
(Rahman et al., 2022). Conventional waste management practices rectify inaccuracies and augment the accuracy of adherence. The
frequently depend on labour-intensive sorting and decision-making incorporation of automation into these models enables the streamlining
procedures, which are recognized for their protracted duration, in of the process, resulting in an enhanced level of accuracy.
efficiency, and susceptibility to human fallibility. The use of GPT has the
potential to improve the existing waste and recycling process. Through 5.5.5. Environmental impact assessment
the utilization of their proficiency in data analysis and pattern identi The demolition process has appreciable environmental kickbacks, in
fication, these models have the capacity to enhance the efficiency of addition to the waste materials generation, the emission of pollutants,
waste sorting, discern recyclable resources, and streamline the and the destruction of ecosystems. Hence, it becomes imperative to
decision-making procedure pertaining to the appropriate disposal of evaluate and address these repercussions to establish and uphold sus
hazardous substances. tainable and ecologically conscientious methodologies (Uzair et al.,
2019).
5.5.3. Redevelopment plan Conventionally, the undertaking of environmental impact assess
Following the dismantling of a physical edifice, a novel and auspi ments and the subsequent implementation of effective management
cious occasion arises, presenting the potential for revitalizing and strategies have been associated with a considerable allocation of human
reutilizing the premises. The formulation of a successful redevelopment resources and temporal commitment. The research endeavors encom
plan presents a formidable undertaking, necessitating meticulous ex pass a comprehensive process of data collection, rigorous analysis, and
amination of a multitude of factors including site conditions, market adherence to regulatory protocols. The evaluation of the environmental
demands, and regulatory stipulations (Volk et al., 2018). ramifications associated with a demolition undertaking presents a
Traditionally, the formulation of these plans is dependent on manual multifaceted endeavor, necessitating the consideration of diverse ele
methodologies that involve comprehensive research, rigorous data ex ments including atmospheric conditions, aquatic contamination,
amination, and extensive engagement with relevant parties. The acoustic disturbances, and ecological diversity. This intricacy com
methods exhibit distinctive attributes including their protracted dura pounds the difficulties encountered by practitioners within the con
tion, substantial resource demands, and proneness to human biases. struction sector.
Furthermore, the complex and intricate nature of incorporating multiple In the present context, it is evident that GPT models present auspi
variables and constraints frequently presents difficulties in ascertaining cious prospects for the optimization and augmentation of environmental
the optimal strategies for redevelopment. To effectively tackle this impact assessment and management procedures. The models possess the
challenge, it is worth noting that GPT models have emerged as a valu capacity to effectively analyse extensive quantities of data derived from
able tool that can provide significant support in the formulation and a wide range of sources, encompassing environmental databases, sci
implementation of well-informed and efficient redevelopment strate entific literature, and historical project data. By employing ML algo
gies. The models possess the capacity to effectively analyse extensive rithms and employing natural language processing techniques, GPT
and heterogeneous datasets, encompassing market trends, demographic models demonstrate the capacity to extract noteworthy insights, identify
data, zoning regulations, and environmental factors. This analytical potential environmental hazards, and suggest appropriate measures for
capability facilitates thorough evaluations of the site’s inherent possi mitigation.
bilities and the discernment of feasible alternatives for redevelopment.
In addition, it is worth noting that GPT models possess the capability to 5.5.6. Establishment of structural issues
effectively support scenario modelling endeavours. This attribute en Before demolition commences, thoroughly evaluating structures is
ables stakeholders to actively engage in the exploration of diverse crucial for identifying issues and ensuring safe, efficient operations.
redevelopment strategies, thereby allowing for a comprehensive evalu However, manual inspection methods are subjective, prone to human
ation of the potential outcomes associated with each strategy. error, and limited in analyzing extensive datasets - potentially missing
concealed defects (Xia et al., 2021), leading to delays, costs, and safety
5.5.4. Regulatory compliance and permitting issues. GPT models have the potential to strengthen structural assess
The strict adherence to regulatory frameworks at various adminis ments by uncovering concealed defects that human visual inspections
trative levels, including local, regional, and national, assumes para may overlook. Examining historical data, maintenance records, and
mount importance in guaranteeing the secure implementation of sensor data can allow GPTs to identify patterns and correlations. This
demolition activities. This adherence serves to effectively mitigate po enables them to highlight sections necessitating additional inspection or
tential risks posed to the environment, public health, and the occupa repair prior to demolition. Rather than depending entirely on subjective
tional safety of workers engaged in such operations. The conventional manual review, combining GPT analysis with current evaluation
methodology commonly involves the laborious implementation of methods could offer more holistic insights into a structure’s
manual procedures to ensure compliance with regulations and obtain vulnerabilities.
necessary permits. The implementation of these procedures typically
necessitates a substantial volume of paperwork, rigorous documentation 5.5.7. Materials recovery planning and maximisation
practices, and intimate cooperation with regulatory entities (Macit İlal Materials recovery and recycling is a significant considerations
and Günaydın, 2017). The circumstances have the potential to lead to within the demolition phase, with immense potential to reduce waste
delays in project timelines and an increase in administrative duties for and environmental impact (Akanbi et al., 2018). Manually developing
15
detailed materials recovery plans is an arduous process requiring sharing and dissemination of knowledge and expertise within an orga
extensive data collection, quantification, and analysis. GPT models can nisation (Saka et al., 2022). There is corpus of construction-related in
augment this process through their ability to quantify recoverable ma formation that were used in the training of GPT models, which enables
terials and identify optimal recovery pathways. By leveraging their NLP these models to be used for knowledge base creation that could entail
and image recognition capabilities (Zheng and Fischer, 2023), these best practices, regulatory requirements, safety procedure, design
models can accurately identify materials such as concrete, steel, wood, guidelines and project documentation. GPT models enable quick and
and plastics (as shown in Fig. 14). This automation streamlines the easy preparation of materials for training the construction personnel on
process of materials identification, enabling efficient sorting and re diverse topics such as health and safety and project management. GPT
covery operations. Another application of GPT in this area is the optimal can be deployed as part of the Chatbot system to provide education
allocation of the recovered material. The models can be integrated with resources for construction personnel based on their queries, thereby,
data on market prices of materials, availability of recycling facilities, providing real-time education, personalized training and improving
and environmental impact assessments. This will enhance the identifi autodidactic experience (Saka et al., 2023).
cation of high-value materials for resale, the determination of suitable GPT models can also be leveraged to capture tacit knowledge that is
recycling options, and the proper disposal of hazardous materials. difficult to manage in the construction industry. The experience and
expertise of construction professionals can be captured and preserved by
5.5.8. Demolition risk assessment using GPT models to gain insights and knowledge from historical data,
Processes such as structural dismantling, debris removal, and haz and past project reports which would enable the creation of knowledge
ardous material management are associated with demolition activities that would benefit new construction professionals. In addition, the
(Akanbi et al., 2018). Conducting a thorough risk assessment is crucial construction industry is a global industry, and the construction site often
to identify potential dangers, prioritize safety measures, and comply faces language barriers. GPT models can provide multilingual support
with regulatory requirements. Traditionally, the risk assessment is per for knowledge management and training. As such, materials, documents
formed by relying on expert judgment, which may be subjective and and information can be translated into different languages in a global
time-consuming (Alipour-Bashary et al., 2022). Furthermore, they may team and for effective dissemination across language boundaries. This is
not fully capture the complex interactions and dependencies between very important as language barriers often limit productivity and hamper
different risk factors. effective health and safety on construction sites.
The emergence of GPT offers a valuable solution to enhance demo
lition risk assessments. GPT models have the potential to analyse various 5.6.2. Customer services
risk factors and provide more accurate and objective assessments by GPT models can be leveraged to improve customer experience and
leveraging their advanced NLP and ML capabilities. The training data satisfaction of firms involved in the delivery of construction projects.
may encompass demolition project records, structural characteristics, GPT models can be used as Chatbot to provide instantaneous assistance
historical accident data, and safety guidelines. This enables the models to customers and can handle requests and provide information about
to identify high-risk areas, predict potential failure modes, and recom project progress, product pricing and other general inquires. GPT
mend appropriate control measures to mitigate risks. models can also be employed to provide tailored automated responses to
customers’ inquiries by identifying the customer’s intent from their
messages. Similarly, GPT models can enable better customer services
5.6. Value-added services
across language boundaries and can also be employed in sentiment
analysis of feedback from customers. Furthermore, GPT can provide
These are opportunities that are not directly related to a specific
suggestions about design and construction based on past data from cli
phase of the construction project lifecycle as shown in Fig. 15.
ents thereby improving personalized interactions and overall customer
services.
5.6.1. Knowledge management and training
Knowledge management and training are perhaps one of the com
5.6.3. Stakeholder communication
mon opportunities for which construction firms can leverage GPT
Construction projects require stakeholders of diverse interests,
models. Knowledge management deals with identifying, capturing,
Fig. 14. Waste management using ChatGPT.
16
Fig. 15. Value-added services.
information needs and power to collaborate for effective delivery. GPT and entity extraction, and natural language generation. As such, these
model can personalise project updates and reports for all the stake models would enable conversational interaction via audio or text with
holders to keep up with project developments and milestones. Similarly, the backup application. For instance, GPT models can be employed for
GPT model can be integrated with the project database and other information retrieval in BIM via natural language (Zhang et al., 2023)
communication platforms to facilitate real-time interaction and and can also be leveraged in the development of extant systems such as
engagement for the stakeholders. Also, GPT models are effective in BIM-based AI voice assistant (Elghaish et al., 2022), Voiced-based vir
summarization and can be leveraged in providing executive summaries tual agents (Linares-Garcia et al., 2022). Leveraging GPT models in
of project-related documentation for effective communication with the conversational system development would improve natural language
stakeholders Lastly, communication between the project team and generation compared to the pattern-based approach currently used in
stakeholders can be analysed by GPT for sentiment analysis and multi some systems which are rigid and susceptible to errors. Also, GPT
lingual communication can be facilitated towards effective stake models would limit data requirements and engineering expertise
holders’ management. necessary for developing prototypes (Zhang et al., 2023) Similarly, with
capabilities of GPT model to generate codes, the model can be leveraged
5.6.4. Market analysis in the development of low code and no code systems. For instance,
Leveraging the advanced data analysis capability of GPT models, platforms can be developed for developing applications for all pro
trend identification and analysis can be done by integrating market data, fessionals in the AEC industry usage without having technical expertise
reports and other relevant data into GPT. Also, insights about market in coding. Drag & drop feature and natural language prompts will be
dynamics, sentiment analysis of customers and market segmentation can employed by the users for interaction with the system, thereby
be carried out to improve products and services. Similarly, they can increasing accessibility and significantly reducing the cost of develop
provide a market analysis of competitors’ services, products and mar ment in the AEC industry.
keting, and customer feedback to gain a competitive advantage and
refine their market strategies. Also, based on the historical data market 6. Challenges
shift, potential market scenarios can be predicted. Lastly, GPT models
can help with summarizing market intelligence reports to support stra The challenges to the deployment of LLMs in the construction in
tegic planning and organisational decision-making, and for monitoring dustry could be broadly classified into three categories – industry-
industry news and updates. related, LLM-related and a nexus of industry and LLM-related as
shown in Fig. 16.
5.6.5. Conversational system/chatbot and low code/no code development
GPT models can be leveraged in the development of conversational
systems/chatbots for natural language recognition, intent classification
17
Fig. 16. Challenges of GPT models in the Construction Industry.
6.1. Hallucinations JSON, CSV, JSONL, TSV, or XLSX) which imply that leveraging con
struction data in GPT models or integrating GPT models with existing
Despite the improvement in NLP in GPT models, the models are tools would involve conversion and pre-processing. For instance, con
prone to hallucination - giving sound and plausible information that is verting BIM model to JSON format, however, not all the information in
not true-which reduces system performance and users’ expectations (Ji the BIM model can be converted which poses a challenge to the
et al., 2023). This is highlighted as a major barrier to the application of deployment of GPT models.
GPT models in the construction industry by the experts because wrong
information could endanger lives and properties and could also lead to
6.3. Domain-specific knowledge and regulatory compliance
profit loss by companies. For instance, relying on information from GPT
models for health and safety management or for programming robotics
Although GPT models are large language models and trained on large
sequences could lead to accidents on site. Similarly, relying on the
data sets, their ability to understand domain-specific knowledge is
hallucinated text on scheduling from GPT models could lead to project
limited. This is a major barrier in the construction industry which re
delay or cost overrun. A typical example of hallucinated text from the
quires a technical understanding of different principles, best practices
GPT model is shown in Fig. 17.
and regulations. As such, there is a need for adequate fine-tuning of GPT
This sounds convincing but it is entirely false as there is no CIDEL
models and the provision of context to improve their performance in a
electric shower from Triton Cara. Rather, CIDEL is a research laboratory
technical domain like the construction industry. Also, Retrieval-
‘Construction Informatics and Digital Enterprise Laboratory’ at Leeds
Augmented Generation (RAG) can be employed to improve the perfor
Beckett University.
mance of the model. Similarly, the regulations in the construction in
dustry are many, as it is one of the most regulated industries, due to
safety and quality needs. These regulations vary from time to time and
6.2. Data and interoperability
are based on different contexts such as countries with different envi
ronmental regulations, labour, and safety-related laws. Most impor
Structured data is needed in the fine-tuning of GPT models which are
tantly, these regulations are technically drafted and require technical
often not readily available in the construction industry. Availability and
knowledge and logical reasoning to understand. Consequently, there is
quality of data have been a major challenge for the application of arti
currently a need for an adequate breakdown of the regulations for easy
ficial intelligence in the industry (Abioye et al., 2021). Also, construc
understanding of GPT models especially complex pictorial regulations
tion projects have unique attributes, and the industry is still slow in the
and the need to update the understanding of the models as regulations
adoption of digital technologies for data collection, leading to potential
change from time to time.
data loss. This is coupled with the heterogeneity of the data and the
time-consuming & labour-intensive exercise in creating labelled data
sets. 6.4. Confidential and intellectual property
Furthermore, the cost of interoperability in the construction industry
is enormous and ranges from millions of pounds (Shehzad et al., 2021). GPT models are trained on large datasets, and these models improve
Datasets are available in different formats such as Portable document their performance with interaction over time. In the construction in
format (PDF), doc, hypertext markup language (HTML), dustry, sensitive data such as project design, cost, contracts, and
Computer-aided design (CAD), Industry foundation classes (IFC) and schedules are generated and could be used as inputs for GPT model
others. Also, different software and digital tools employed have different applications. This raises the concern of confidentiality of GPT models
formats that make seamless interoperability difficult. Data are often generating output with sensitive information. Also, using data such as
converted to a standardized form like IFC to improve interoperability project designs and patented building techniques could infringe on the
but at a loss of watering down the richness of the data (Saka et al., 2023). intellectual property owner and lead to ethical and legal issues. Lastly,
GPT models currently only accept specific formats (Image formats, the ownership of data has always been a problem in the construction
18
Fig. 17. Hallucinated outputs.
industry and the use of GPT models begs the question of who can use the fear that AI would take over jobs in the construction industry and the
data created during the construction project for training large language perceived complexity of digital tools. Consequently, trust and accept
models. Consequently, there is a need for a clear policy and guidelines ability are inherent challenges in the construction industry that would
for the use of GPT models in the construction industry to avoid confi hinder the applications of GPT models. There is a need to demonstrate
dentiality and intellectual property challenges. the tangibility of GPT models and highlight their reliability through
robust testing and validation. It is important to employ an inclusive
6.5. Trust and acceptability approach which would involve all stakeholders in considering the usage
of GPT models to have clear guidelines and to allay fears.
The construction industry is known to be slow in the adoption of
innovation compared to other industries. There is often resistance to 6.6. Liability and ethics
change in the industry, and digital technologies such as BIM have not
been widely implemented as expected. With the growing application of GPT can be leveraged for many opportunities at the design, con
AI in the construction industry, industry practitioners and stakeholders struction, management, and demolition stages of construction projects.
are sceptical about trusting and accepting it. Decisions by stakeholders However, a growing area of concern in the use of AI like GPT models is
have significant implications for cost, safety, quality and time of project liability and accountability challenges (McAleenan, 2020). GPT models
delivery and relying on GPT models that are ‘black box’ is difficult. As are LLM trained on a large amount of data which influences what the
projects often involve different parties, the issue of acceptance by some models would generate during deployment. Bias, incomplete informa
parties would hinder the deployment of GPT models on such projects tion and inaccuracies in the training data would affect the output
and could also influence the unavailability of project data to fine-tune generated by the models, and this could cause harm and business loss in
the models. Furthermore, LLMs do require access to large databases the construction industry. Similarly, the inability of GPT models to fully
for training, with construction firms having concerns about the data comprehend various legal and regulatory requirements in the industry
usage, and are often unwilling to release such data. This is coupled with could lead to the provision of inaccurate and non-compliant advice and
19
recommendations from the system. Additionally, the decision-making (OpenAI, 2023a,b). Gpt-3.5-turbo (ChatGPT) employed for use case
mechanisms within GPT models present challenges in interpretability, validation cost $0.002/1K tokens (OpenAI, 2023a,b). Consequently,
transparency, and explainability. Such obscurity undermines stake aside from the infrastructure requirement, the cost of leveraging LLMs
holder trust and complicates accountability efforts. Individuals or en are barrier to deployment in the construction industry. Chen et al.
tities depending on the outputs or recommendations generated by these (2023) proposed strategies to leverage LLMs at a reduced cost with
models often find themselves in a position of uncertainty regarding the improved performance. The study highlighted prompt adaption, LLM
internal algorithms or systematic processes utilized. (You et al., 2023). approximation, and LLM cascade as strategies; and validated ‘Frugal
These beg the question of who is responsible for the potential harm GPT’ (LLM cascade) with 98% cost reduction.
caused by wrong information generated from GPT models. There are
currently no clear legal frameworks on accountability and liability for 6.9. Scalability and performance optimization
the harm caused by the application of AI in the construction industry,
which could lead to exploitation and leave affected parties with no clear Whilst GPT models have overcome some of the scalability challenges
legal recourse. However, the optimal current approach to overcome li inherent in the development of Conversational AI systems (Saka et al.,
ability and accountability challenges is to have clear policies and 2023) via a generic model that can be personalized with a prompt
guidelines for the usage of GPT models, human oversight and to employ design, the challenge of scalability of fine-tuned models persists.
models as complement to the human personnel. Fine-tuning improves learning of GPT models by training on more
Similarly, the problem of ethical issues facing LLM affects GPT structured training data for a specific use case. This could be classifi
models in the construction industry. GPT models have the capability of cation (e.g sentiment analysis, email categorization) or conditional
amplifying biases and discrimination learnt from the training data set, generation (e.g entity extraction, chatbot for customer support)
can be used unethically, and could impact the labour market in the (OpenAI, 2023a,b) and would require different datasets. However, such
construction industry. As such, it is important, for the GPT models to be specialization necessitates the use of disparate datasets and conse
trained on unbiased and diverse datasets to ensure accurate and reliable quently limits the model’s scalability across diverse tasks. For instance, a
results. During the deployment in construction organisations, it is model fine-tuned for customer support chatbot functionalities would
important to have clear guidelines to prevent the model’s misuse. In likely underperform in tasks related to information retrieval from
addition, there are economic impacts of GPT models in the industry Building Information Modeling (BIM). Similarly, while the performance
which could lead to the loss of jobs, pose as a competitor, and replace of GPT models can be optimized by better prompt designs, other ap
repetitive tasks such as information retrieval from texts done by some proaches such as self-reflection (Reflexion) require skilled personnel to
low-skilled workers in the industry. It is thus important to take steps to improve the models (Shinn et al., 2023). Also, fine-tuning models to
evaluate the impact of GPT models in the AEC industry and mitigate optimize performance requires structured data and infrastructure,
negative effects. which might be difficult for the majority of the organisations in the
construction industry.
6.7. Skills and training
6.10. Cybersecurity
Effective deployment of GPT models in the construction industry
requires new skillsets and training programs to ensure that the pro In an era of digital advancements, cybersecurity has become a crit
fessionals can leverage them properly. The most needed skill is ‘prompt ical concern across various industries, including the construction sector
engineering’ which deals with developing and optimizing prompts to (Nyamuchiwa et al., 2022). As the construction industry embraces the
effectively use LLMs (Liu et al., 2023). It is encompassing skills needed to potentials of GPT, it also faces significant cybersecurity challenges.
build, interact, and develop new capabilities with LLMs. There are The utilization of GPT within construction processes brings forth
different techniques for prompt engineering such as zero-shot, few-shot fresh areas of vulnerability that can be manipulated by malicious in
and chain-of-thought (CoT) prompting (Wei et al., 2022). In zero-shot dividuals. The dependency on interconnected systems, cloud computing,
prompting, no examples are provided for the models to perform speci and data exchange amplifies the likelihood of unauthorized entry, data
fied tasks whilst in few-shot prompting provide contexts and examples in breaches, and cyber-attacks (You and Feng, 2020). It is imperative for
the prompts for the model to perform better (Wei et al., 2022). On the construction firms to confront these obstacles in order to protect sensi
other hand, CoT prompting provides a ‘series of intermediate reasoning tive information and uphold the steadfastness of their operations.
steps’ which significantly boost the ability of LLMs to perform complex Alongside external cyber threats, construction companies must also
reasoning tasks. Further, there is a need to understand the fine-tuning address insider threats. These can arise from unintentional errors or
process of GPT models and the use of RAG which enables the models deliberate actions by employees who have access to sensitive informa
to perform better than prompt design (OpenAI, 2023a,b). However, this tion (Nyamuchiwa et al., 2022). Proper employee training and aware
requires structured training examples in specific formats (JSON, CSV, ness programs are crucial to mitigate the risk of internal data breaches or
JSONL, TSV, or XLSX) which implies the need to have the skillsets to compromises. It is vital for organisations to foster a
pre-process data. As such, skills, and training required for the proper cybersecurity-conscious culture and promote best practices among em
deployment of GPT models in the construction industry serves as ployees. Furthermore, construction companies need to invest in
bottlenecks. advanced threat detection systems, intrusion detection and prevention
systems, and real-time monitoring tools. Prompt response to security
6.8. Infrastructure requirement and cost incidents is crucial to minimize potential damage and prevent unau
thorized access to critical systems or data.
Leveraging LLMs in the construction industry would require neces
sary infrastructure such as computing power, network connectivity and 6.11. Interdisciplinary collaboration
data storage. This poses a challenge for the small and medium-sized
enterprises which represent about 80% of the organisation in the in In the dynamic and complex landscape of the construction industry,
dustry (Saka and Chan, 2020). Also, the current cost of web access for successful project outcomes often hinge on effective collaboration
ChatGPT is $20/month with restricted traffic. Similarly, the develop among diverse disciplines (Oraee et al., 2019). Interdisciplinary
ment of applications that leverage GPT models is not free, as usage is collaboration involves the integration of expertise from various fields,
billed per 1000 tokens (~750 words). This ranges from $0.09/1K to such as architecture, engineering, construction management, and data
$0.18/1K (prompt and completion cost) for 8K and 32K GPT-4 model science, to address the multifaceted challenges encountered throughout
20
the project lifecycle (Fulford and Standing, 2014). processes. To address this challenge, continuous model training and
While the introduction of GPT in the construction industry brings adaptation are crucial.
promising opportunities, it also presents unique challenges to achieving The integration of GPT into construction processes involves a com
seamless interdisciplinary collaboration. Professionals from different plex technical ecosystem comprising servers, databases, network infra
disciplines often use specialized language and terminology specific to structure, and data processing systems. Coordinating and managing
their respective fields. GPT-generated outputs may include technical these components can be challenging, particularly as the scale and
terms and jargon that might be unfamiliar to professionals from other complexity of GPT applications increase. This challenge can be miti
disciplines. This can lead to misunderstandings, misinterpretations, and gated by putting in place dedicated IT personnel responsible for infra
inefficiencies in collaborative efforts. For successful interdisciplinary structure management, regular system health checks, proactive
collaboration, trust and acceptance among team members are crucial. troubleshooting, and prompt response to technical issues.
The introduction of GPT may raise concerns and scepticism among
professionals who are unfamiliar with its capabilities or who fear the
6.15. Multilingual language processing
potential displacement of their roles. Building trust and fostering
acceptance of GPT as a collaborative tool can be a significant challenge,
In an increasingly globalized world, the construction industry is
requiring effective change management strategies (Sezgin et al., 2022).
experiencing a growing need for multilingual language processing ca
pabilities. With projects spanning across borders and involving diverse
6.12. Cultural and social considerations
stakeholders, effective communication and understanding of multiple
languages have become crucial (Kraft, 2019; Taiwo et al., 2022).
Construction projects are diverse and often take place in various
Although GPT models are constantly being updated with more training
cultural and social contexts, involving different stakeholders with
datasets (Brown et al., 2020), models trained in limited languages may
distinct values, norms, and practices (Zuo et al., 2012). These cultural
struggle to accurately comprehend and generate content in unfamiliar or
and social factors can significantly influence the acceptance, adoption,
less-represented languages. This variability poses a significant hurdle in
and effectiveness of GPT-driven solutions.
achieving seamless multilingual communication.
Due to their automation capabilities, the implementation of GPT in
Using zero-shot learning, Lai et al. (2023) conducted extensive ex
the construction industry may raise ethical and social considerations
periments on ChatGPT (i.e., powered by GPT 3.5) to investigate its
about the impact on employment and job security. The fear of job
multilingual capacity using various NLP tasks such as summarization,
displacement and the potential loss of skilled labour can create resis
question answering, named entity recognition, and part-of-speech
tance to adopting emerging technologies such as GPT models (Na et al.,
tagging, amongst others. Their result indicated that ChatGPT under
2022). Moreover, concerns regarding data privacy, ownership, and se
performed in various languages except for English compared to
curity can undermine trust and hinder the widespread acceptance of
task-specific models built in specific languages. This is due to the fact
GPT-empowered interventions. In order to mitigate these challenges, the
that a larger percentage of ChatGPT training data was in English (Kou
relevant stakeholders can implement reskilling and upskilling programs
baa et al., 2023). To address the challenge of language variability, it is
to equip the workforce with new skills required to work alongside GPT
essential to create and curate large-scale multilingual training datasets
systems and clearly communicate data handling practices to ensure
that encompass a wide range of languages, dialects, and domains spe
compliance with relevant regulations.
cific to the construction industry. By incorporating diverse linguistic
data, GPT models can better understand and generate content in various
6.13. Latency issue
languages, enhancing their multilingual language processing
capabilities.
The latency issue relates to the time lag that occurs between the input
of data and the output of results (Yang et al., 2023). In the construction
industry, latency can be especially problematic because of the 6.16. Standards and variability
time-sensitive nature of construction projects.
The GPT algorithms require to be fine-tuned on massive amounts of The construction industry involves diverse standards and re
data to achieve their highest level of performance (Zheng and Fischer, quirements that vary significantly across countries, regions, and indi
2023). Hence, the training process can be time-consuming. An addi vidual projects (Cheriyan and Choi, 2020). This poses a challenge in
tional factor contributing to latency stems from the insufficiency of efficiently leveraging GPT models across the entire lifecycle of con
processing capabilities in certain systems. GPT algorithms demand struction projects. Strategies to address this challenge include adaptive
considerable computational resources, necessitating high-performance databases, customized fine-tuning, prompt conditioning,
hardware for expedited processing. The financial burden associated cost-effectiveness analysis, time management considerations and
with procuring such hardware may be unfeasible for smaller construc multilingual models. Firstly, instead of relying on a fixed corpus, GPT
tion enterprises, thereby exacerbating latency challenges. Such delays models can be trained on adaptive databases that are regularly updated
can adversely affect various operations, including real-time monitoring with the latest standards, codes, and project specifications from different
of construction sites, collaborative design efforts, and the functioning of regions. This allows the models to account for geographical variations
automated robotic systems. and changes over time. The databases can ingest structured data from
BIM, project documents, and compliance repositories (Bilal et al., 2016).
6.14. Maintenance and upkeep Secondly, GPT models pre-trained on broad construction data can be
fine-tuned on project-specific documents and asset data to adapt to the
While GPT systems offer numerous benefits, ensuring their contin requirements of individual projects. This customization improves the
uous operation and optimal performance over time requires diligent relevance of the models’ outputs for each unique project context.
maintenance practices and regular updates (Zong and Krishnamachari, Similarly, Retrieval-Augmented Generation which provides a frame
2022). One of the primary challenges in GPT maintenance is system work for improving LLM-generated responses by grounding LLMs on
degradation and drift. Over time, GPT models may experience a decline external verifiable information can be employed for the diverse stan
in performance due to changing data patterns, evolving user re dards and requirements. Lastly, Training multilingual GPT models on
quirements, or shifting industry standards. This degradation can lead to construction data from diverse regions can aid in overcoming language
reduced accuracy, reliability, and relevance of generated outputs, barriers and differences in terminology across geographies. This expands
impacting the overall effectiveness of GPT applications in construction the applicability across international projects Lai et al. (2023).
21
7. Case study validation: material selection and optimization reduce the level of randomness in the model’s responses.
platform The prompt consists of two parts: the user prompt and the system
prompt. The system prompt is concatenated with the properties of the
To validate the practical application of GPT in the construction in element in context which is extracted from BIM JSON search before
dustry, we developed a material selection and optimization prototype. being sent to the ChatGPT API. User prompts serve as the initial input
This prototype aimed to validate a use case for GPT in the design phase. and should clearly express the user’s intention, such as requesting
To facilitate this, a simple BIM model containing 235 building elements assistance in finding the best material for a specific element in the BIM
was employed as a test case. The development process involves three key model. Providing sufficient context in prompts helps guide the conver
steps, each of which serves a specific purpose. Fig. 18 shows the system sation effectively (Fig. 20). System instructions play a crucial role in
architecture for the use case validation. guiding the behaviour and responses of the ChatGPT. They leveraged the
output from the BIM JSON search through the data retrieval module to
7.1. Data retrieval module refine prompts. For example, system instructions may direct ChatGPT to
consider factors, such as moisture resistance, when suggesting materials
The Forge Model Derivative API is used to translate and process the for a bathroom door located at the entrance. These instructions, com
BIM models. The BIM file is uploaded to the forge cloud using the Data bined with the user prompts and BIM data search results, ensure that
Management API and stored for access through the Model Derivative ChatGPT generates relevant and informed recommendations.
API (Fig. 19). The model was then translated into the SVF2 format, Prompt development is an iterative process that incorporates zero-
enabling the extraction of geometric and metadata information. The shot and few-shot scenarios. The performance of the model is continu
Forge Viewer renders 2D and 3D models for web-based access to BIM ally evaluated, and the prompts, instructions, and dialogue system are
data. The Model Derivative API extracts data attributes and converts refined based on user feedback and real-world interactions. This itera
them to a searchable JSON format, including hierarchy trees, geome tive approach enhanced the efficiency and effectiveness of the system
tries, and properties such as component ID, type, and area location. The over time.
Data Retrieval module effectively extracts and cleans the necessary BIM
data, making it useable for subsequent modules, such as NLP. 7.3. User interface and integration module
7.2. NLP prompt processing module The user interface (UI) development for the material selection and
optimization system involved the use of JavaScript, HTML, CSS3, and
The prompt development for material selection and optimization Bootstrap 5 (Fig. 21). These technologies were utilized to create a chat
using the ChatGPT API is a multistep process that involves constructing interface that seamlessly integrates with the forge viewer, which pro
appropriate prompts and designing an interactive dialogue system. The vides a library for visualizing cloud-based BIM. To facilitate smooth
goal is to guide users in selecting optimal materials for each element of communication between the front-end interface and the prompt man
the BIM model based on their requirements. The prompt manager con ager, an AJAX call was implemented. This allowed for the exchange of
nects to the OpenAI API with the ChatGPT model 3.5 turbo using AJAX time-stamped natural language queries and answers as well as the
asynchronous communication. A temperature value of 0.5 is set to retrieval of result IDs. Using these IDs, cloud-based BIM rendered
Fig. 18. System architecture.
22
Fig. 19. Extraction of element property from model derivative API.
Fig. 20. Few-shot system prompt instruction.
different 3D contextual scenes of the relevant building objects in 7.4.1. Zero-shot prompting
response to the user’s query, enabling interactive exploration of the To test the zero-shot capabilities, the prototype disabled the system
uploaded BIM model. role prompt in the payload of the ChatGPT API call, retaining only the
The web-based prototype was designed to be easily deployed to any user prompt and supplying BIM information as input.
cloud service provider, ensuring high availability and accessibility from Fig. 17 shows the response generated by chatGPT in response to a
various devices, including mobile devices. The UI seamlessly integrates user query “Suggest the best material to be used for the door named ‘Bois
BIM data available through the OpenAI API, providing a comprehensive - Panneau de porteto’ leading to toilet WC 15” at zero-shot scenario.
and user-friendly experience for material selection and optimization. ChatGPT’s response to the query was not specific as it responded
with “As an AI language model, I am not aware of the specific requirements
7.4. Discussion of the door named "Bois - Panneau de porteto" leading to toilet WC 15” as
demonstrated in Fig. 8 above. However, it suggested common material
Fig. 22 showcases the developed prototype for material selection and for bathroom while laying emphasis on the specific requirements of the
optimization. The user interface provides seamless integration of user user and then recommending consulting professionals to determine the
input queries and ChatGPT responses, both displayed on the right side of best material for the user’s specific needs. The above scenario demon
the interface. Additionally, the prototype incorporates a visual repre strated that ChatGPT could not serve as an effective tool for material
sentation of the 3D model using the Autodesk Forge viewer. selection and optimization without proper prompt engineering.
The prototype was tested in 3 different scenarios zero-shot, few-shot
with system prompting and edge case prompting scenarios. 7.4.2. Few-shot prompting scenario
Fig. 23 shows the response generated by ChatGPT in a few-shot
scenario response to the same user query used in the zero-shot
23
Fig. 21. Interface set-up and Integration to ChatGPT.
scenario above with the user query: "Suggest the best material to be used query, conducted through the data retrieval module, confirms that the
for the door to toilet WC 15." ChatGPT successfully provides optimized wall in question is an external wall before calling up the ChatGPT API
material recommendations for a bathroom door, taking into consider (GPT 3.5 Turbo model), which then analyses possible materials, taking
ation various factors. In contrast to zero-shot scenario, it carefully avoids into account the available data within the model, such as the external
suggesting wood as it is prone to moisture damage in bathroom settings nature of the wall and the user’s requirement for energy efficiency to
and acknowledges the privacy concerns associated with glass, although suggest an insulated concrete wall with an R-value of up to 20.
glazed glass could provide a privacy solution. Instead, ChatGPT rec The few-shot scenario showcases the remarkable capability of
ommends steel and composite materials as viable options, owing to their ChatGPT as a valuable tool for guiding home builders in selecting the
moisture resistance and durability. most suitable materials for their constructions. By leveraging both
ChatGPT demonstrates its understanding of a specific door element general information and user-defined preference parameters, ChatGPT
and its location within the BIM model. It recognizes that a door to a offers informed recommendations that align with the desired charac
bathroom, specifically WC15, is susceptible to water damage owing to teristics of building components.
its proximity to moisture sources. Based on the available data from the
BIM model, ChatGPT analyses possible materials for the door, compares 7.4.3. Edge case prompting scenario
them using factors such as location, area of the component, and other In Fig. 25, when ChatGPT was prompted with a question ‘What is
relevant attributes. Furthermore, ChatGPT considers additional factors BIM’ that was not related to material selection and optimization, it
such as material availability, cost, and energy efficiency during the se responded with the message, "I apologize, but I am unable to handle
lection process. The user also has the option to specify the factor that your request as it is not related to material selection and optimization".
holds the utmost importance, enabling ChatGPT to prioritize that factor This indicates that ChatGPT can be tailored to specific needs and
in its final recommendation. deployed as an expert system in a particular field. The prototype allows
Fig. 24 displays another few-shot scenario where the user query for easy leveraging of the extensive language capabilities of ChatGPT by
“suggests the energy efficient wall named Maconnerie Isolant.” The BIM readjusting the prompts, without the need for technical experts as
24
Fig. 22. Zero-shot scenario.
Fig. 23. Few-shot scenario.
required by other NLP systems. This capability empowers construction These enhancements will enhance the overall performance and reli
practitioners to actively participate in rapid testing and prototyping ability of the material selection and optimization system.
processes, making the system more accessible and user-friendly for in
dustry professionals. 8. Conclusion
7.5. Limitations Application of LLMs/Generative AI such as GPT models is still low

despite these models overcoming some of the previous challenges that
The prototype demonstrated the capability of selecting the best make the deployment of AI difficult in the construction industry. For
material for a specific component in a BIM model. However, it requires instance, the time and expertise needed in the development of Conver
the identification of the component to provide accurate responses. sational AI systems have been drastically reduced by the GPT models
Future studies would include enhancing the system by probing compo trained on large databases. Sectors such as education, medicine, and
nents and their properties in multiple conversations and leveraging the business have been leveraging these new large language models to
seamless NLP feature of ChatGPT. This involves guiding the user with a improve their modus of Operandi. However, little is known about the
list of parameters to choose from, such as cost, geographical location, opportunities and limitations of leveraging GPT models in the con
and aesthetics, to improve the selection process. Engaging industry struction industry. Consequently, this study employed three sequential
stakeholders allows for the collection of query and response data, which steps to identify and evaluate the opportunities and limitations of GPT
can be used to fine-tune the ChatGPT model for better efficiency. In models in the construction industry. A detailed preliminary study was
addition, the prompt development process will be evaluated and conducted and reinforced with an expert discussion on the opportunities
improved to minimize incorrect or irrelevant responses to user queries. and limitations of GPT models. Based on the identified opportunities, a
25
Fig. 24. Prompting (Happy path user).
Fig. 25. Prompting (Edge case).
use case was validated, and a prototype was developed for material thoughts learning via prompt designs or integration with existing sys
selection and optimization by integrating BIM with a GPT model. tems. Similarly, some of the opportunities require the need to fine-tune
The study reveals that although large language model such as BERT GPT models with structured data to improve the performance of the
has been gaining attention in the construction industry, GPT models are models and leverage existing or external databases (RAG). However,
not well known in the industry until the recent launch of ChatGPT. The despite these immense opportunities, the study shows that the deploy
study highlights that current applications of GPT models in the industry ment of GPT models in the construction industry would need to over
are for information retrieval, scheduling, and logistics. Opportunities come challenges that are inherent in the GPT models and within the
identified for the deployment of GPT models span the predesign, design, construction industry. Inherent limitations of GPT models such as hal
construction, and post-construction phases of the project. Also, value- lucinations, accepted input formats, cost and reliability were high
added services are discussed as part of the opportunities for GPT lighted in the study. On the other hand, challenges such as trust,
models which include areas that are not directly related to a specific acceptability, domain technicalities, skills, and interoperability which
phase of the project lifecycle. These opportunities were discussed, and ensue because of the construction industry context were revealed.
the findings reveal that the GPT models can be leveraged for some of the Furthermore, the study validated a use case with one of the opportu
opportunities with zero-shot learning, few-shot learning, chain-of- nities – materials selection and optimization – by integrating BIM and
26
GPT. The data retrieval module, NLP prompt processing module, user problem. Expert Syst. Appl. 198 (February), 116728 https://doi.org/10.1016/j.
eswa.2022.116728.
interface and integration module were developed for the prototype. The
Amaral, R.E.C., Brito, J., Buckman, M., Drake, E., Ilatova, E., Rice, P., Sabbagh, C.,
study shows that leveraging the developed prototype would improve Voronkin, S., Abraham, Y.S., 2020. Waste management and operational energy for
efficiency by providing the stakeholders with materials and optimizing sustainable buildings: A review. Sustainability 12 (13). https://doi.org/10.3390/
the selection based on defined objectives such as the location of the su12135337.
Amer, F., Jung, Y., Golparvar-Fard, M., 2021. Transformer machine learning language
components and cost. model for auto-alignment of long-term and short-term plans in construction. Autom.
In addition, there are limitations that serve as fertile grounds for ConStruct. 132 https://doi.org/10.1016/j.autcon.2021.103929.
further research. The preliminary search for literature was conducted in As, I., Pal, S., Basu, P., 2018. Artificial intelligence in architecture: generating conceptual
design via deep learning. Int. J. Architect. Comput. 16 (4), 306–327. https://doi.org/
Scopus, google scholar, web of science and validated in other databases 10.1177/1478077118800982.
(ACM and Science direct) which could serve as a limitation. Similarly, Asiedu, R.O., Adaku, E., Owusu-Manu, D.-G., 2017. Beyond the causes. Construct.
the search query developed, and the inclusion criteria of the English Innovat. 17 (3), 363–380. https://doi.org/10.1108/ci-01-2016-0003.
Assaf, M., Hussein, M., Abdelkhalek, S., Zayed, T., 2023. A multi-criteria decision-
language could have led to the missing of some publications. Also, the making model for selecting the best project delivery systems for offsite construction
size of the expert recruited for the discussion could serve as a limitation; projects. Buildings 13 (2), 571. https://doi.org/10.3390/BUILDINGS13020571/S1.
however, due diligence was observed to ensure the experts have the Bach, C., Bernhaupt, R., D’Agostini, C.S., Winckler, M., 2013. Mobile applications for
incident reporting systems in urban contexts: lessons learned from an empirical
right experience. The developed prototype was also not subjected to any study. ACM International Conference Proceeding Series. https://doi.org/10.1145/
quantitative validation to evaluate its performance. Further studies can 2501907.2501960.
explore the different opportunities areas highlighted in this study by Bavarian, M., Jun, H., Tezak, N., Schulman, J., McLeavey, C., Tworek, J., Chen, M., 2022.
Efficient Training of Language Models to Fill in the Middle. http://arxiv.org/abs/22
developing prompts, fine-tuning and integrating GPT models with
07.14255.
existing databases and systems. Also, an advanced extension of the Beach, T.H., Hippolyte, J., Rezgui, Y., 2020. Towards the adoption of automated
materials selection and optimization system is currently being devel regulatory compliance checking in the built environment. Autom. ConStruct. 118
oped by the researchers. This study contributes to the growing body of (May), 103285 https://doi.org/10.1016/j.autcon.2020.103285.
Beata, P.A., Jeffers, A.E., Kamat, V.R., 2018. Real-time fire monitoring and visualization
knowledge on the application of GPT models and provides research for the post-ignition fire state in a building. Fire Technol. 54 (4), 995–1027. https://
vistas for leveraging GPT models in the construction industry. It high doi.org/10.1007/s10694-018-0723-1.
lights the current limitations of LLMs and the need to overcome these Bilal, M., Oyedele, L.O., Qadir, J., Munir, K., Ajayi, S.O., Akinade, O.O., Owolabi, H.A.,
Alaka, H.A., Pasha, M., 2016. Big Data in the construction industry: A review of
barriers for the proliferation of LLMs in the construction industry. Its present status, opportunities, and future trends. Adv. Eng. Inf. 30 (3), 500–521.
findings would be of benefit to researchers and stakeholders in the AEC Boehm, B.W., 1984. Verifying and validating software requirements and design
industry. specifications. IEEE Soft. 1 (1), 75–88. https://doi.org/10.1109/MS.1984.233702.
Bouabdallaoui, Y., Lafhaj, Z., Yim, P., Ducoulombier, L., Bennadji, B., 2021. Predictive
maintenance in building facilities: A machine learning-based approach. Sensors 21
(4), 1–15. https://doi.org/10.3390/s21041044.
Declaration of competing interest Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A.,
Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G.,
Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., et al., 2020.
The authors declare that they have no known competing financial
Language models are few-shot learners. Advances in Neural Information Processing
interests or personal relationships that could have appeared to influence Systems 33, 1877–1901.
the work reported in this paper. Budayan, C., Dikmen, I., Birgonul, M.T., 2015. Alignment of project management with
business strategy in construction: evidence from the Turkish contractors. J. Civ. Eng.
Manag. 21 (1), 94–106. https://doi.org/10.3846/13923730.2013.802737.
Data availability Castro-lacouture, D., Asce, A.M., Süer, G.A., Gonzalez-joaqui, J., Yates, J.K., 2009.
Construction project scheduling with time , cost. Mater. Restrictions Using Fuzzy
Data will be made available on request. Math. Mod. Critical Path Method 135, 1096–1104. https://doi.org/10.1061/(ASCE)
0733-9364(2009)135. October.
Catenda, 2020. Employer’s Information Requirements (EIR). https://catenda.com/glo
Acknowledgements ssary/bim-eir/.
Chaphalkar, N.B., Iyer, K.C., Patil, S.K., 2015. Prediction of outcome of construction
dispute claims using multilayer perceptron neural network model. Int. J. Proj.
The authors would like to acknowledge and express their sincere Manag. 33 (8), 1827–1835. https://doi.org/10.1016/j.ijproman.2015.09.002.
gratitude to Leeds Beckett University for providing financial support for Chen, Y., Kamara, J.M., 2005. The Use of Mobile Computing in Construction Information
this study. Management. 21st Annual Conference of the Association of Researchers in
Construction Management. ARCOM) SOAS, London.
Chen, L., Zaharia, M., Zou, J., 2023. FrugalGPT: How to Use Large Language Models
References while Reducing Cost and Improving Performance arXiv preprint arXiv:2305.05176.
Cheng, M.Y., Tsai, H.C., Hsieh, W.S., 2009. Web-based conceptual cost estimates for
construction projects using Evolutionary Fuzzy Neural Inference Model. Autom.
Abioye, S.O., Oyedele, L.O., Akanbi, L., Ajayi, A., Davila Delgado, J.M., Bilal, M.,
ConStruct. 18 (2), 164–172. https://doi.org/10.1016/j.autcon.2008.07.001.
Akinade, O.O., Ahmed, A., 2021. Artificial intelligence in the construction industry:
Cheriyan, D., Choi, J.-H., 2020. A review of research on particulate matter pollution in
A review of present status, opportunities and future challenges. J. Build. Eng. 44
the construction industry. J. Clean. Prod. 254, 120077 https://doi.org/10.1016/j.
https://doi.org/10.1016/j.jobe.2021.103299.
jclepro.2020.120077.
Abu Bakar, N.N., Hassan, M.Y., Abdullah, H., Rahman, H.A., Abdullah, M.P., Hussin, F.,
Chew, K.C., 2010. Singapore’s strategies towards sustainable construction. IES J. Part A
Bandi, M., 2015. Energy efficiency index as an indicator for measuring building
Civ. Struct. Eng. 3 (3), 196–202. https://doi.org/10.1080/19373260.2010.491641.
energy performance: A review. Renew. Sustain. Energy Rev. 44, 1–11. https://doi.
Cornwell, N., Bilson, C., Gepp, A., Stern, S., Vanstone, B.J., 2022. The Role of Data
org/10.1016/j.rser.2014.12.018.
Analytics within Operational Risk Management: A Systematic Review from the
Adedara, M.L., Taiwo, R., Bork, H.-R., 2023. Municipal solid waste collection and
Financial Services and Energy Sectors, vol. 74, pp. 374–402. https://doi.org/
coverage rates in sub-saharan african countries: A comprehensive systematic review
10.1080/01605682.2022.2041373, 1.
and meta-Analysis. Waste 1, 389–413.
Dannoun, Y., 2022. Application of supply chains management in construction project : a
Akanbi, L.A., Oyedele, L.O., Akinade, O.O., Ajayi, A.O., Davila Delgado, M., Bilal, M.,
review in the compatibility between the procurements and implementation process.
Bello, S.A., 2018. Salvaging building materials in a circular economy: A BIM-based
Int. J. Adv. Eng., Sci. Appl. 3 (1), 18–21.
whole-life performance estimator. Resour. Conserv. Recycl. 129, 175–186. https://
Decoder, 2022. OpenAI Cuts Prices for GPT-3 by Two Thirds. https://the-decoder.com/
doi.org/10.1016/j.resconrec.2017.10.026. May 2017.
openai-cuts-prices-for-gpt-3-by-two-thirds/.
Akinosho, T.D., Oyedele, L.O., Bilal, M., Ajayi, A.O., Delgado, M.D., Akinade, O.O.,
Devlin, J., Chang, M.W., Lee, K., Toutanova, K., 2018. BERT: pre-training of deep
Ahmed, A.A., 2020. Deep learning in the construction industry: A review of present
bidirectional transformers for language understanding. In: NAACL HLT 2019 - 2019
status and future innovations. J. Build. Eng. 32 https://doi.org/10.1016/j.
Conference of the North American Chapter of the Association for Computational
jobe.2020.101827.
Linguistics: Human Language Technologies - Proceedings of the Conference,
Alipour-Bashary, M., Ravanshadnia, M., Abbasianjahromi, H., Asnaashari, E., 2022.
pp. 4171–4186, 1. https://arxiv.org/abs/1810.04805v2.
Building demolition risk assessment by applying a hybrid fuzzy FTA and fuzzy
Dikbas, A., Ergen, E., Giritli, H., 2010. A review of the artificial intelligence applications
CRITIC-TOPSIS framework. Int. J. Build. Pathol. Adapt. 40 (1), 134–159. https://doi.
in construction dispute resolution. In: Managing IT in Construction/Managing
org/10.1108/IJBPA-08-2020-0063.
Construction for Tomorrow. Routledge.
Al-shihabi, S., Mladenović, N., 2022. A mixed integer linear programming model and a
basic variable neighbourhood search algorithm for the repatriation scheduling
27
Dimyadi, J., Pauwels, P., Spearpoint, M., Clifton, C., Amor, R., 2015. Querying a Kotei, E., Thirunavukarasu, R., 2023. A systematic review of transformer-based pre-
regulatory model for compliant building design Audit. In: Proc. Of the 32nd CIB W78 trained Language Models through self-supervised learning. Information 2023 14 (3),
Conference 2015, pp. 139–148. https://doi.org/10.13140/RG.2.1.4022.6003, 27th- 187. https://doi.org/10.3390/INFO14030187, 14, Page 187.
29th October 2015, Eindhoven, The Netherlands, October. Koubaa, A., Boulila, W., Ghouti, L., Alzahem, A., Latif, S., 2023. Exploring ChatGPT
Dolphin, W.S.Y., Alshami, A.A.M., Tariq, S., Boadu, V., Mohandes, S.R., Ridwan, T., capabilities and limitations : A critical review of the NLP game changer. Preprints,
Zayed, T., 2021. Effectiveness of policies and difficulties in improving safety March, 1–18. https://doi.org/10.20944/preprints202303.0438.v1.
performance of repair, maintenance, minor alteration, and addition works in Hong Kraft, K., 2019. Language policies and linguistic competence: new speakers in the
Kong. Int. J. Construct. Manag. 0 (0), 1–30. https://doi.org/10.1080/ Norwegian construction industry. Lang. Pol. 18 (4), 573–591. https://doi.org/
15623599.2021.1935130. 10.1007/s10993-018-9502-6.
Egan, J., 1998. Rethinking Construction: Report of the Construction Task Force on the Kulkarni, P., Mahabaleshwarkar, A., Kulkarni, M., Sirsikar, N., Gadgil, K., 2019.
Scope for Improving the Quality and Efficiency of UK Construction. Department of Conversational AI: an overview of methodologies, applications & future scope. 5th
the Environment & R. (DETR). T. a. t. https://constructingexcellence.org.uk/wp-co International Conference on Computing, Communication, Control and Automation
ntent/uploads/2014/10/rethinking_construction_report.pdf. (ICCUBEA). IEEE, 2019 1–7. https://doi.org/10.1109/
Elghaish, F., Chauhan, J.K., Matarneh, S., Pour Rahimian, F., Hosseini, M.R., 2022. ICCUBEA47591.2019.9129347.
Artificial intelligence-based voice assistant for BIM data management. Autom. Kusimo, H., Oyedele, L., Akinade, O., Oyedele, A., Abioye, S., Agboola, A., Mohammed-
ConStruct. 140 https://doi.org/10.1016/j.autcon.2022.104320. Yakub, N., 2019. Optimisation of resource management in construction projects: a
El-Omari, S., Moselhi, O., 2009. Integrating automated data acquisition technologies for big data approach. World J. Sci. Technol. Sustain. Dev. 16 (2), 82–93. https://doi.
progress reporting of construction projects. 2009 26th International Symposium on org/10.1108/wjstsd-05-2018-0044.
Automation and Robotics in Construction. ISARC 2009 86–94. https://doi.org/ Lai, V.D., Ngo, N.T., Veyseh, A. P. Ben, Man, H., Dernoncourt, F., Bui, T., Nguyen, T.H.,
10.22260/isarc2009/0048. 2023. ChatGPT beyond English: towards a comprehensive evaluation of large
Emovon, I., Oghenenyerovwho, O.S., 2020. Application of MCDM method in material Language Models in multilingual learning. http://arxiv.org/abs/2304.05613.
selection for optimal design: A review. Res. Mater. 7 https://doi.org/10.1016/j. Levy, S.M., 2010. Construction process planning and management - CHP5- preparing the
rinma.2020.100115. bid documents. Construct. Process Plann. Manag. 113–135.
Ezelogs, 2023. Ultimate & Exciting Decision Making Using Chat GPT in Construction. Lewis, R., Séquin, C., 1998. Generation of 3D building models from 2D architectural
https://blogs.ezelogs.com/decision-making-using-chat-gpt/. plans. CAD Comput. Aided Des. 30 (10), 765–779. https://doi.org/10.1016/S0010-
Firat, M., 2023. How ChatGPT Can Transform Autodidactic Experiences and Open 4485(98)00031-1.
Education? Preprint. https://doi.org/10.31219/osf.io/9ge8m. Li, J., Dada, A., Kleesiek, J., Egger, J., 2023. ChatGPT in Healthcare: A Taxonomy and
Florez, L., Castro-Lacouture, D., 2013. Optimization model for sustainable materials Systematic Review. medRxiv. https://doi.org/10.1101/2023.03.30.23287899.
selection using objective and subjective factors. Mater. Des. 46, 310–321. https:// Linares-Garcia, D.A., Roofigari-Esfahan, N., Pratt, K., Jeon, M., 2022. Voice-based
doi.org/10.1016/j.matdes.2012.10.013. intelligent virtual Agents (VIVA) to support construction worker productivity.
Fulford, R., Standing, C., 2014. Construction industry productivity and the potential for Autom. ConStruct. 143 https://doi.org/10.1016/j.autcon.2022.104554.
collaborative practice. Int. J. Proj. Manag. 32 (2), 315–326. https://doi.org/ Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G., 2023. Pre-train, prompt, and
10.1016/j.ijproman.2013.05.007. predict: A systematic survey of prompting methods in Natural Language processing.
Gambatese, J.A., Hallowell, M., 2011. Factors that influence the development and ACM Comput. Surv. 55 (9), 1–35. https://doi.org/10.1145/3560815.
diffusion of technical innovations in the construction industry. Construct. Manag. Liu, X., Jing, X., Zhu, Q., Du, W., Wang, X., 2023. Automatic construction hazard
Econ. 29 (5), 507–517. https://doi.org/10.1080/01446193.2011.570355. identification integrating on-site scene graphs with information extraction in
Grussing, M.N., 2014. Life cycle Asset management methodologies for buildings. outfield test. Buildings 13 (2), 377. https://doi.org/10.3390/buildings13020377.
J. Infrastruct. Syst. 20 (1), 1–8. https://doi.org/10.1061/(asce)is.1943- Lu, Y., Sood, T., Chang, R., Liao, L., 2020. Factors Impacting Integrated Design Process of
555x.0000157. Net Zero Energy Buildings: an Integrated Framework, vol. 22, pp. 1700–1712.
Globerson, S., Zwikael, O., 2002. Impact of the project manager on project management https://doi.org/10.1080/15623599.2020.1742625, 9.
planning processes. Proj. Manag. J. 31 (3), 58–64. https://doi.org/10.1177/ Macit İlal, S., Günaydın, H.M., 2017. Computer representation of building codes for
875697280203300308. automated compliance checking. Autom. ConStruct. 82, 43–58. https://doi.org/
Gozalo-Brizuela, R., Garrido-Merchan, E.C., 2023. ChatGPT Is Not All You Need. A State 10.1016/j.autcon.2017.06.018. May 2016.
of the Art Review of Large Generative AI Models arXiv preprint arXiv:2301.04655. Mantel, 2023. ChatGPT Decoded: A Comprehensive Overview of Large Language Models.
Guo, K., Zhang, L., 2022. Multi-objective optimization for improved project Marjaba, G.E., Chidiac, S.E., 2016. Sustainability and resiliency metrics for buildings -
management: current status and future directions. Autom. ConStruct. 139, 104256 critical review. Build. Environ. 101, 116–125. https://doi.org/10.1016/j.
https://doi.org/10.1016/J.AUTCON.2022.104256. buildenv.2016.03.002.
Hagras, H., Packham, I., Vanderstockt, Y., McNulty, N., Vadher, A., Doctor, F., 2008. An McAleenan, P., 2020. Moral responsibility and action in the use of artificial intelligence
intelligent Agent based Approach for energy management in commercial buildings. in construction. Proc. Inst. Civil Eng.- Manag. Procurement Law 173 (4), 166–174.
Fuzzy Syst. Conf. 156–162. https://doi.org/10.1680/jmapl.19.00056.
Hayman, P., 2022. Chat GPT Is Revolutionizing the World of Architecture and Design. Mcbride, K., Noordt, C. Van, Misuraca, G., Hammerschmid, G., 2021. Towards a
Wooduchoose. https://www.wooduchoose.com/blog/ai-and-architecture/. Systematic Understanding on the Challenges of Procuring Artificial Intelligence in
Hernández, A., Amigó, J.M., 2021. Attention mechanisms and their Applications to the Public Sector. September.
complex systems. Entropy 2021 23 (3), 283. https://doi.org/10.3390/E23030283, Miller, S., 2016. An aggregated Approach to risk Analysis: risk portfolios. Risk portfolios.
23, Page 283. Enterprise Risk Manag.: Common Framework Entire Organ. 141–149. https://doi.
Hon, C.K.H., Chan, A.P.C., Chan, D.W.M., 2011. Strategies for improving safety org/10.1016/B978-0-12-800633-7.00010-9.
performance of repair, maintenance, minor alteration and addition (RMAA) works. Mills, A., 2001. A systematic approach to risk management for construction. Struct. Surv.
Facilities 29 (13/14), 591–610. https://doi.org/10.1108/02632771111178391. 19 (5), 245–252. https://doi.org/10.1108/02630800110412615/FULL/XML.
Hsu, C., Sandford, B.A., 2007. The Delphi technique: making sense of consensus. Momade, M.H., Durdyev, S., Estrella, D., Ismail, S., 2021. Systematic review of
Practical Assess. Res. Eval. 12 (10), 1–8. application of artificial intelligence tools in architectural, engineering and
Hussin, F., Md Rahim, S.A.N., Hatta, N.S.M., Aroua, M.K., Mazari, S.A., 2023. construction. Front. Eng. Built Environ. 1 (2), 203–216. https://doi.org/10.1108/
A systematic review of machine learning approaches in carbon capture applications. FEBE-07-2021-0036.
J. CO2 Util. 71. https://doi.org/10.1016/J.JCOU.2023.102474. MUO, 2023. GPT-1 to GPT-4: Each of OpenAI’s GPT Models Explained and Compared.
Ibrahim, Y.M., Lukins, T.C., Zhang, X., Trucco, E., Kaka, A.P., 2009. Towards automated https://www.makeuseof.com/gpt-models-explained-and-compared/.
progress assessment of workpackage components in construction projects using Na, S., Heo, S., Han, S., Shin, Y., Roh, Y., 2022. Acceptance model of Artificial
computer vision. Adv. Eng. Inf. 23 (1), 93–103. https://doi.org/10.1016/j. intelligence (AI)-Based technologies in construction firms: Applying the technology
aei.2008.07.002. Acceptance model (TAM) in combination with the
Jenkins, D.A., Smith, T.E., 1994. Applying Delphi methodology in family therapy technology–organisation–environment (TOE) framework. Buildings 12 (2). https://
research. Contemp. Fam. Ther. 16 (5), 411–430. https://doi.org/10.1007/ doi.org/10.3390/buildings12020090.
bf02197902. Neelakantan, A., Xu, T., Puri, R., Radford, A., Han, J.M., Tworek, J., Yuan, Q., Tezak, N.,
Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y.J., Madotto, A., Fung, P., Kim, J.W., Hallacy, C., Heidecke, J., Shyam, P., Power, B., Nekoul, T.E., Sastry, G.,
2023. Survey of hallucination in Natural Language generation. ACM Comput. Surv. Krueger, G., Schnurr, D., Such, F.P., Hsu, K., et al., 2022. Text and Code Embeddings
55 (12), 1–38. https://doi.org/10.1145/3571730. by Contrastive Pre-training. http://arxiv.org/abs/2201.10005.
Jiang, H., 2019. Mobile fire evacuation system for large public buildings based on Nyamuchiwa, K., Lei, Z., Aranas, C., 2022. Cybersecurity vulnerabilities in off-site
Artificial intelligence and IoT. IEEE Access 7, 64101–64109. https://doi.org/ construction. Appl. Sci. 12 (10), 1–25. https://doi.org/10.3390/app12105037.
10.1109/ACCESS.2019.2915241. Oluleye, B.I., Chan, D.W.M., Saka, A.B., Olawumi, T.O., 2022. Circular economy research
Karhade, M., 2022. OpenAI Released GPT-3 Text-Davinci-003. I Compared it with 002. on building construction and demolition waste: A review of current trends and future
The Results Are Impressive! https://pub.towardsai.net/openai-just-released-gpt-3-te research directions. J. Clean. Prod. 357 https://doi.org/10.1016/j.
xt-davinci-003-i-compared-it-with-002-the-results-are-impressive-dced9aed0cba. jclepro.2022.131927.
Kim, C., Son, H., Kim, C., 2013. Automated construction progress measurement using a OpenAI, 2019. GPT-2: 1.5B Release. OpenAI Blog.
4D building information model and 3D data. Autom. ConStruct. 31, 75–82. https:// OpenAI, 2020. OpenAI Licenses GPT-3 Technology to Microsoft. OpenAI Blog.
doi.org/10.1016/j.autcon.2012.11.041. https://openai.com/blog/openai-licenses-gpt-3-technology-to-microsoft/.
Kim, Y., Chin, S., Choo, S., 2022. BIM data requirements for 2D deliverables in OpenAI, 2022. Introducing ChatGPT. OpenAI.Com. https://openai.com/blog/chatgpt.
construction documentation. Autom. ConStruct. 140, 104340 https://doi.org/ OpenAI, 2023a. Fine-tuning Retrieved 2nd of May 2023 from. https://platform.openai.co
10.1016/J.AUTCON.2022.104340. m/docs/guides/fine-tuning/fine-tuning.
OpenAI, 2023b. GPT-4 Technical Report. 4, 1–100. http://arxiv.org/abs/2303.08774.
28
Oraee, M., Hosseini, M.R., Edwards, D.J., Li, H., Papadonikolaki, E., Cao, D., 2019. Taiwo, R., Shaban, I.A., Zayed, T., 2023a. Development of sustainable water
Collaboration barriers in BIM-based construction networks: A conceptual model. Int. infrastructure: A proper understanding of water pipe failure. J. Clean. Prod. https://
J. Proj. Manag. 37 (6), 839–854. https://doi.org/10.1016/j.ijproman.2019.05.004. doi.org/10.1016/j.jclepro.2023.
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Tezel, A., Koskela, L., 2023. Off-site construction in highways projects: management,
Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., technical, and technology perspectives from the United Kingdom. https://doi.org/1
Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., Lowe, R., 2022. 0.1080/01446193.2023.2167218.
Training Language Models to Follow Instructions with Human Feedback. htt Tixier, A.J.P., Hallowell, M.R., Rajagopalan, B., Bowman, D., 2016. Application of
p://arxiv.org/abs/2203.02155. machine learning to construction injury prediction. Autom. ConStruct. 69, 102–114.
Paaß, G., Giesselbach, S., 2023. Foundation Models for Natural Language Processing – https://doi.org/10.1016/j.autcon.2016.05.016.
Pre-trained Language Models Integrating Media. https://arxiv.org/abs/2302.0 Togal, A.I., 2023. Harnesses ChatGPT for Construction. https://www.forconstructionpro
8575v1. s.com/construction-technology/article/22820209/togalai-togalai-harnesses-chatgpt
Parm, A.G., 2023. How ChatGPT Can Help in Project Management. https://parm. -for-construction.
com/en/chatgpt-in-project-management/. Uddin, S.J., Albert, A., Ovid, A., Alsharef, A., 2023. Leveraging ChatGPT to Aid
Porter, S., 2021. Construction Industry Looking to Predictive Analytics to Improve construction hazard recognition and support safety education and training.
Project Outcomes. https://www.oracle.com/news/announcement/idc-construction- Sustainability 15 (9), 7121.
industry-looks-to-predictive-analytics-to-improve-project-outcomes-2021-10-27/. Uzair, M., Chun, D., Han, H., Jeon, G., Chen, K., 2019. Energy & Buildings A review of
Prieto, S.A., Mengiste, E.T., García de Soto, B., 2023. Investigating the use of ChatGPT for the applications of artificial intelligence and big data to buildings for energy-
the scheduling of construction projects. Buildings 13 (4). https://doi.org/10.3390/ efficiency and a comfortable indoor living environment. Energy Build. 202, 109383
buildings13040857. https://doi.org/10.1016/j.enbuild.2019.109383.
Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I., 2022. Robust Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł.,
Speech Recognition via Large-Scale Weak Supervision. http://arxiv.org/abs/2212 Polosukhin, I., 2017. Attention Is All You Need 31st Conference on Neural
.04356. Information Processing Systems (NIPS 2017). Long Beach, CA, USA.
Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., 2018. Improving Language VentureBeat, 2020. OpenAI Makes GPT-3 Generally Available through its API. https
understanding by generative pre-training. Homol. Homotopy Appl. 9 (1), 399–438. ://venturebeat.com/2021/11/18/openai-makes-gpt-3-generally-available-through
https://www.cs.ubc.ca/~amuham01/LING530/papers/radford2018improving.pdf. -its-api/?fbclid=IwAR0Ox7ymcC_pg88e9YEC-KmjHUPx-_DWwanwB7IbLcKbQ3tNw
Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., 2019. Language Models t6fXBQAU8Y.
Are Unsupervised Multitask Learners. Volk, R., Luu, T.H., Mueller-Roemer, J.S., Sevilmis, N., Schultmann, F., 2018.
Rahman, M.W., Islam, R., Hasan, A., Bithi, N.I., Hasan, M.M., Rahman, M.M., 2022. Deconstruction project planning of existing buildings based on automated
Intelligent waste management system using deep learning with IoT. J. King Saud acquisition and reconstruction of building information. Autom. ConStruct. 91,
University - Comput. Inform. 34 (5), 2072–2087. https://doi.org/10.1016/j. 226–245. https://doi.org/10.1016/j.autcon.2018.03.017. July 2017.
jksuci.2020.08.016. Wang, Y.R., Yu, C.Y., Chan, H.H., 2012. Predicting construction cost and schedule
Ratnasabapathy, S., Rameezdeen, R., 2010. A decision support system for the selection of success using artificial neural networks ensemble and support vector machines
best procurement system in construction. Built-Environ. Sri Lanka 7 (2), 43. https:// classification models. Int. J. Proj. Manag. 30 (4), 470–478. https://doi.org/10.1016/
doi.org/10.4038/BESL.V7I2.1943. j.ijproman.2011.09.002.
Ribeirinho, M.J., Mischke, J., Strube, G., Sjödin, E., Blanco, J.L., Palter, R., Biörck, J., Wei, C., Wang, Y.-C., Wang, B., Kuo, C.-C.J., 2023. An Overview on Language Models:
Rockhill, D., Andersson, T., 2020. The Next Normal in Construction: How Disruption Recent Developments and Outlook. https://arxiv.org/abs/2303.05759v1.
Is Reshaping the World’s Largest Ecosystem. M. Company. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Chi, E., Le, Q., Zhou, D., 2022. Chain-of-
Saka, A.B., Chan, D.W.M., 2019. Knowledge, skills and functionalities requirements for Thought prompting elicits reasoning in large Language Models. arXiv preprint.
quantity surveyors in building information modelling (BIM) work environment: an https://doi.org/10.48550/arXiv.2201.11903.
international Delphi study. Architect. Eng. Des. Manag. 16 (3), 227–246. https://doi. Xia, Y., Lei, X., Wang, P., Sun, L., 2021. Artificial intelligence based structural assessment
org/10.1080/17452007.2019.1651247. for regional short-and medium-span concrete beam bridges with inspection
Saka, A.B., Chan, D.W.M., 2020. Profound barriers to building information modelling information. Rem. Sens. 13 (18) https://doi.org/10.3390/rs13183687.
(BIM) adoption in construction small and medium-sized enterprises (SMEs). Yang, J., Jin, H., Tang, R., Han, X., Feng, Q., Jiang, H., Yin, B., Hu, X., 2023. Harnessing
Construct. Innovat.: Inf. Process. Manag. 20 (2), 261–284. https://doi.org/10.1108/ the Power of LLMs in Practice: A Survey on ChatGPT and beyond, vol. 1, pp. 1–24, 1.
ci-09-2019-0087. http://arxiv.org/abs/2304.13712.
Saka, A.B., Chan, D.W.M., Wuni, I.Y., 2022. Knowledge-based decision support for BIM Yeung, H.C., Ridwan, T., Tariq, S., Zayed, T., 2020. BEAM Plus implementation in Hong
adoption by small and medium-sized enterprises in developing economies. Autom. Kong : assessment of challenges and policies BEAM Plus implementation in Hong
ConStruct. 141 https://doi.org/10.1016/j.autcon.2022.104407. Kong : assessment of challenges and policies. Int. J. Construct. Manag. 0 (0), 1–15.
Saka, A.B., Oyedele, L.O., Akanbi, L.A., Ganiyu, S.A., Chan, D.W., Bello, S.A., 2023. https://doi.org/10.1080/15623599.2020.1827692.
Conversational artificial intelligence in the AEC industry: A review of present status, You, H., Ye, Y., Zhou, T., Zhu, Q., Du, J., 2023. Robot-enabled construction assembly
challenges and opportunities. Adv. Eng. Inf. 55, 101869 https://doi.org/10.1016/j. with Automated sequence planning based on ChatGPT: RoboGPT. arXiv preprint.
aei.2022.101869. https://doi.org/10.48550/arXiv.2304.11018.
Sezgin, E., Sirrianni, J., Linwood, S.L., 2022. Operationalizing and implementing You, Z., Feng, L., 2020. Integration of industry 4.0 related technologies in construction
pretrained, large Artificial intelligence linguistic models in the US health care industry: A framework of cyber-physical system. IEEE Access 8, 122908–122922.
system: outlook of generative pretrained transformer 3 (GPT-3) as a service model. https://doi.org/10.1109/ACCESS.2020.3007206.
JMIR Med. Inform. 10 (2), 1–7. https://doi.org/10.2196/32875. Zhang, H., Song, H., Li, S., Zhou, M., Song, D., 2022. A survey of controllable text
Shaalan, K., 2010. Rule-based Approach in Arabic Natural Language processing. generation using transformer-based pre-trained Language Models. J. ACM 37 (4). htt
International Journal on Information and Communication Technologies (IJICT). htt p://arxiv.org/abs/2201.05337.
p://www.ieee.ma/IJICT/IJICT-SI-Bouzoubaa-3.3/3-Khaledl.pdf. Zhang, J., El-Gohary, N., 2016. Automated information transformation for Automated
Shehzad, H.M.F., Ibrahim, R.B., Yusof, A.F., Khaidzir, K.A.M., Iqbal, M., Razzaq, S., 2021. regulatory compliance checking in construction. J. Comput. Civ. Eng. https://doi.
The role of interoperability dimensions in building information modelling. Comput. org/10.1061/(ASCE)CP.1943-5487.0000427. November.
Ind. 129 https://doi.org/10.1016/j.compind.2021.103444. Zhang, S., Fan, R., Liu, Y., Chen, S., Liu, Q., Zeng, W., 2023. Applications of transformer-
Shinn, N., Labash, B., Gopinath, A., 2023. Reflexion: an autonomous agent with dynamic based language models in bioinformatics: a survey. Bioinform Adv 3 (1), vbad001.
memory and self-reflection. arXiv preprint. https://doi.org/10.48550/ https://doi.org/10.1093/bioadv/vbad001.
arXiv.2303.11366. Zheng, J., Fischer, M., 2023. BIM-GPT: a prompt-based virtual Assistant framework for
Siraj, N.B., Asce, S.M., Fayek, A.R., Ph, D., Eng, P., Asce, M., 2019. Risk Identification BIM information retrieval. arXiv preprint. https://doi.org/10.48550/arXiv.2304.
and Common Risks in Construction: Literature Review and Content Analysis, vol. 09333.
145. https://doi.org/10.1061/(ASCE)CO.1943-7862.0001685, 9. Zong, M., Krishnamachari, B., 2022. A Survey on GPT-3. ArXiv. http://arxiv.org/abs/22
Taiwo, R., Ben Seghier, M.E.A., Zayed, T., 2023. Towards sustainable water 12.00857.
infrastructure : the state-of-the-art for modeling the failure probability of water Zuo, J., Zillante, G., Coffey, V., 2012. Project culture in the Chinese construction
pipes. Water Resour. Res. 59, e2022WR033256 https://doi.org/10.1029/ industry: perceptions of contractors. Australasian J. Construct. Econom. Build. 9 (2),
2022WR033256. 17–28.
Taiwo, R., Hussein, M., Zayed, T., 2022. An integrated Approach of simulation and
regression Analysis for Assessing productivity in modular integrated construction
projects. Buildings 12. https://doi.org/10.3390/buildings12112018.
29

Developments in The Built Environment

Uploaded by

Copyright:

Available Formats

Developments in The Built Environment

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Developments in The Built Environment

Uploaded by

Copyright:

Available Formats

Developments in the Built Environment 17 (2024) 100300

Contents lists available at ScienceDirect

Developments in the Built Environment

GPT models in construction industry: Opportunities, limitations, and a use

1. Introduction advancement in information technologies and digital tools, the AEC

Fig. 1. Structure of the paper.

Fig. 2. Research approach.

Fig. 3. Opportunities for GPT models in the Pre-design Phase.

Fig. 4. Act-As prompting.

Fig. 5. Conversational development of Client’s brief.

Fig. 6. Opportunities for GPT models in the Design Phase.

Fig. 8. Opportunities for GPT models in Construction Phase.

Fig. 9. Hazard identification using computer vision capability of ChatGPT.

5.4.4. Lifecycle management of asset 5.4.6. Regulatory compliance management

Fig. 11. Facility chatbot.

Fig. 12. Compliance report generation.

Fig. 13. Opportunities for GPT models in the Demolition Phase.

Fig. 14. Waste management using ChatGPT.

Fig. 15. Value-added services.

Fig. 16. Challenges of GPT models in the Construction Industry.

Fig. 17. Hallucinated outputs.

Fig. 18. System architecture.

Fig. 19. Extraction of element property from model derivative API.

Fig. 20. Few-shot system prompt instruction.

Fig. 21. Interface set-up and Integration to ChatGPT.

Fig. 22. Zero-shot scenario.

Fig. 23. Few-shot scenario.

7.5. Limitations Application of LLMs/Generative AI such as GPT models is still low

Fig. 24. Prompting (Happy path user).

Fig. 25. Prompting (Edge case).

You might also like