Latest / 1 More Paper / Prioritising generative AI adoption challenges in construction projects: a Fuzzy Analytic Hierarchy Process approach — Fernández & Shen (2026)
Transcript
- You’re listening to “Prioritising generative AI adoption challenges in construction projects: a Fuzzy Analytic Hierarchy Process approach,” by A. Vidal Fernández and Y. Shen. Published in 2026. ISSN: 1562-3599 (Print) 2331-2327 (Online) Journal homepage: the linked source Prioritising generative AI adoption challenges in construction projects: a Fuzzy Analytic Hierarchy Process approach Alejandro Vidal Fernández & Yixue Shen To cite this article: Alejandro Vidal Fernández & Yixue Shen (07 May 2026): Prioritising generative AI adoption challenges in construction projects: a Fuzzy Analytic Hierarchy Process approach, International Journal of Construction Management, DOI: 10.1080/15623599.2026.2664476 © 2026 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group Published online: 07 May 2026. Submit your article to this journal Article views: 642 View related articles View Crossmark data Prioritising generative AI adoption challenges in construction projects: a Fuzzy Analytic Hierarchy Process approach Alejandro Vidal Fernandez and Yixue Shen School of Civil Engineering, University of Leeds, Leeds, UK ABSTRACT. The construction industry has experienced slow productivity growth over the past decade, and generative artificial intelligence (AI) has emerged as a potential solution. However, there is limited understanding of its adoption challenges, particularly from a decision-oriented prioritisation perspective to support effective implementation. This study aims to identify and prioritise the most significant challenges associated with the adoption of generative AI in construction projects. A systematic literature review (SLR) was conducted to identify the main challenges and sub-challenges discussed in the existing body of literature. Based on the SLR findings, an online questionnaire was developed to collect expert judgements from industry practitioners. The data were analysed using the Fuzzy Analytic Hierarchy Process (FAHP), and the robustness of the results was examined through sensitivity analysis by simulating variations in criteria importance. The results identify data privacy, security and legal liability; poor quality and availability of training data; and lack of construction-specific knowledge and language as the most critical challenges. These findings underscore the importance of data governance, data quality, and domain-specific capabilities for effective implementation. They also provide practical guidance for managerial decision-making by prioritising critical challenges for the effective adoption of generative AI in construction projects. Introduction. ARTICLE HISTORY Generative artificial intelligence; construction projects; construction productivity; adoption challenges; Fuzzy Analytic Hierarchy Process; decision support Despite the transformative potential of digitalization and automation, the construction industry continues to exhibit limited innovation and, more critically, slow adoption of digital technologies. This limited adoption constrains the effective implementation of emerging technologies and is reflected in the sector’s persistently low productivity growth. Compared with manufacturing, retail, and telecommunications, construction remains among the least productive sectors, with annual productivity growth estimated at approximately 0.8%, equivalent to only one-third of the global economy’s average. This underperformance has substantial economic implications, as the construction sector accounts for approximately 7–10% of GDP in developed economies and 3–6% in developing countries. Addressing this problem is critical to achieving the United Nations Sustainable Development Goals (SDGs), specifically SDG 8 (Decent work and economic growth) and SDG 9 (Industry, innovation and infrastructure). With global construction output projected to increase by 85% by 2030, improving industry performance to levels comparable with more advanced sectors could generate annual savings of up to USD 1.6 trillion, highlighting the importance of overcoming challenges that hinder the effective adoption of emerging technologies in construction projects. Artificial Intelligence (AI) has been recognised as a key enabler for improving project performance across the construction lifecycle. Within this landscape, generative AI, a subset of AI technologies, has attracted increasing attention due to the significant opportunities it offers across the construction sector. However, as an emerging technology, its adoption remains in an early stage and is associated with various challenges across technical, contextual, organisational and ethical dimensions. Several studies emphasise the capacity of generative AI to streamline workflows, automate labour-intensive tasks, and support data-driven decision-making in construction projects. Empirical evidence reported by Isah and Kim (2025), Shi et al. (2025), and Taiwo et al. (2025) further supports these claims, highlighting successful applications in task automation, trend prediction, and strategic planning. Conversely, other studies adopt a more cautious perspective, noting that the deployment of generative AI raises unresolved concerns related to data transparency, algorithmic bias, and information reliability, which require careful consideration prior to large-scale implementation. Despite the growing interest in generative AI within the construction industry, several gaps remain in the existing literature. Prior studies highlight the need to better understand the current state, applications, opportunities, and challenges associated with generative AI adoption across the construction project lifecycle. Although there is a general consensus regarding the types of challenges affecting adoption, particularly those related to large language models such as ChatGPT, most studies address these challenges in isolation, without considering their interaction across different implementation contexts. Moreover, the literature reveals a lack of empirical validation in real-world project settings and under varying conditions, as well as the absence of a structured framework to prioritise which challenges are most critical to address. This limits the practical applicability of existing findings and constrains informed decision-making regarding generative AI adoption. Addressing this gap is essential to support effective implementation strategies and improve overall project outcomes. In response to the gaps identified in the literature, this study aims to integrate a systematic literature review (SLR) with expert-based empirical analysis using the Fuzzy Analytic Hierarchy Process (FAHP) to identify and prioritise the most significant challenges associated with the adoption of generative AI in construction projects. By doing so, the study offers a decision-support perspective that assists construction practitioners and organisations in managing the adoption of generative AI. To achieve this aim, the study pursues the following objectives: I. To systematically identify and classify the key challenges and sub-challenges associated with generative AI adoption in construction projects through a SLR. II. To prioritise the identified challenges based on expert judgement using the FAHP method. III. To evaluate the robustness of the prioritisation results through sensitivity analysis. Literature review. This section synthesises prior research on generative AI adoption challenges in construction projects, drawing on a SLR conducted in accordance with PRISMA guidelines. A detailed description of the SLR protocol is provided in Appendix A. Existing studies identify a range of challenges associated with the deployment of generative AI applications in construction projects. Although these challenges are often discussed in relation to specific use cases, the literature shows that they can be systematically classified into four main categories: domain-specific, technological, adoption-related, and ethical challenges. Domain-specific challenges Domain-specific challenges refer to the difficulties generative AI applications face in addressing construction-specific codes, standards, and specialised engineering knowledge. Samsami (2024) and Smetana et al. (2024) further highlight that the evolving nature of construction regulations requires continuous model fine-tuning, which poses additional challenges for stable implementation. Li et al. (2024) note that many generative AI tools currently applied in construction are trained primarily on general-domain data, leading to limited construction-specific knowledge learning. The frequent need for retraining due to updates in construction standards increases both training costs and implementation time, while also affecting model stability. Technological challenges Technological challenges encompass the technical limitations encountered when integrating generative AI tools into existing construction systems. Prior studies identify issues such as hallucinations, reliability concerns, and limited transparency, particularly in the application of large language models (LLMs). Lee et al. (2024) report that inaccurate outputs or hallucinations have been observed in several LLM-based applications used in construction contexts. In addition, Li et al. (2025) and Taiwo et al. (2025) associate transparency limitations with the black-box nature of these models, which complicates error diagnosis and reduces user trust, especially in high-pressure project environments. Adoption challenges Adoption challenges refer to barriers that restrict the actual use of generative AI tools within organisations. Samsami (2024) and Taiwo et al. (2025) highlight that high implementation costs and limited internal expertise significantly hinder adoption. These costs are commonly attributed to data preparation, model development, increased computational requirements, and system integration. While large construction firms may be able to absorb such investments, medium-sized and small contractors often face prohibitive costs, particularly when return on investment is uncertain (Aladag 2023; Samsami 2024; Taiwo et al. 2025). Furthermore, Heo and Na (2025) and Smetana et al. (2024) emphasise that generative AI implementation requires specialised skills that are frequently lacking within construction organisations. Uddin et al. (2025) further identify difficulties related to recruiting AI-skilled professionals, developing appropriate training pathways, and bridging the gap between AI expertise and construction process knowledge. Ethical challenges Ethical challenges relate to the moral and societal risks associated with the use of generative AI tools. Sonkor and Garcıa de Soto (2025) and Araya-Aliaga et al. (2025) emphasise concerns related to data privacy, potential misuse, and workforce implications as critical considerations during implementation. Lekan et al. (2020) and Li et al. (2024) note that data privacy risks arise from the possibility of unauthorised access to sensitive project information and intellectual property. Increased digital interconnection further heightens exposure to data breaches. In addition, Smetana et al. (2024) and Maceika et al. (2024) highlight the risk of misuse, noting that generative AI tools can produce highly convincing documents, such as budgets and plans, that may be exploited for fraudulent purposes. Concerns related to workforce displacement and task automation may also generate resistance to adoption within project teams Based on the SLR, a set of challenges associated with the adoption of generative AI in construction projects was identified. An initial list of main challenges and their corresponding sub-challenges, together with the literature sources in which they were discussed, was extracted from the reviewed studies. The identified sub-challenges were classified into four categories: domain-specific, technological, adoption, and ethical. As presented in Table 1, the identified challenges and sub-challenges were subsequently consolidated and grouped based on conceptual similarity to ensure clarity and maintain a manageable structure for expert evaluation. This hierarchical classification informed the design of the FAHP questionnaire and enabled the prioritisation of sub-challenges according to their relative importance for the successful implementation of generative AI in construction projects. Materials and methods. This study adopts an analytical decision-support approach to identify and prioritise generative AI adoption challenges in construction projects. The research design consists of four sequential phases, consistent with prior studies published in the International Journal of Construction Management that have employed similar decision-support and multi-criteria approaches. Phase 1 involved conducting a SLR to synthesise existing academic knowledge and identify generative AI adoption challenges in construction projects. Phase 2 focused on the development of a questionnaire based on the challenges identified through the SLR. Phase 3 involved analysing the collected expert judgement data using the FAHP to prioritise the identified challenges according to their perceived criticality. Finally, Phase 4 consisted of conducting a sensitivity analysis to assess the robustness and reliability of the prioritisation results. Fuzzy Analytic Hierarchy Process FAHP was employed to prioritise the identified generative AI adoption sub-challenges based on expert judgement. The method uses pairwise comparisons to evaluate the relative importance of sub-challenges Note: Intermediate values were allowed. within each challenge category and derive priority weights for decision-making. By incorporating triangular fuzzy numbers (TFN), FAHP addresses key limitations of the traditional analytic hierarchy process, particularly the difficulty of representing uncertainty and vagueness in human judgment when assigning crisp numerical values. Triangular fuzzy numbers (1, 3, 5, 9) were used to construct the comparison matrices. Each fuzzy number was represented by three parameters (l, m, u), corresponding to the lower bound, the modal value, and the upper bound, respectively. Table 2 presents the equivalences between the linguistic scale applied in the survey and the corresponding TFN. The prioritisation of sub-challenges based on the pairwise comparison results followed the standard FAHP procedure outlined by Bakhtari et al. (2021) and recommended by Wang et al. (2008). Step 1: Construct pairwise matrices. Pairwise comparison matrices were constructed for each participant across the four sub-challenge categories. The comparisons were expressed using TFN based on a predefined linguistic scale. For each participant, four matrices were generated, corresponding to the different challenge categories. The general structure of each matrix is illustrated in Equation. The matrix notation D, T, A, and E represent domain-specific, technological, adoption-related, and ethical challenges, respectively. The subscript associated with each matrix indicates the participant who provided the pairwise comparisons, while n denotes the number of sub-challenges within each category. Step 2: Calculate the fuzzy geometric mean For each pairwise comparison matrix, the fuzzy geometric mean values were computed TFNs, as expressed in Equation. !1 Y n n r~ Sij i 1⁄4 j1⁄41 Where i and j denote the row and column indices of the pairwise comparison matrix, respectively. Step 3: Calculate the fuzzy weights The fuzzy weights (wi) for each sub-challenge were calculated by normalising each geometric mean parameter (l, m, u), with respect to the sum of all geometric means, as expressed in Equation. !−1 X n wi 1⁄4 ri ri i1⁄41 Where ri denotes the fuzzy geometric mean of each sub-challenge, while the normalisation was performed separately for the lower, modal, and upper bounds (l, m, u) to obtain the corresponding fuzzy weight components. Step 4: Defuzzify and normalise weights The fuzzy weights were defuzzified to derive crisp weights for each sub-challenge. Defuzzification was performed using Equation. Mi 1⁄4 lwi þ mwi þ uwi 3 The resulting crisp weights were subsequently normalised to obtain relative importance values expressed as percentages, as shown in Equation. Mi Pn Ni 1⁄4 i1⁄41 Mi Consistency was assessed using the consistency ratio (CR) proposed by Saaty (1990). Matrices exceeding the accepted threshold were corrected using the logarithmic least squares method (LLSM). The matrices were then aggregated within each challenge category using the fuzzy geometric mean across all participants. Based on the normalised weights (Ni), the sub-challenges were ranked within their respective challenge categories to identify the most critical barriers to the adoption of generative AI in construction projects. All pairwise comparisons were conducted using a common evaluation criterion, namely the perceived impact of each sub-challenge on the successful adoption of generative AI. As all fuzzy weights were derived and normalised following the same FAHP procedure, the resulting weights allow for consistent comparison across different challenge categories. Questionnaire design and expert data collection The questionnaire designed to collect the expert judgements comprised two sections. The first section gathered general information about the participants, including region of practice, educational background, current job position, and years of professional experience. The second section consisted of pairwise comparison questions using a slider-based format to capture expert judgements, with the structure adapted from Wang et al. (2008). A total of 24 experts participated in the survey. This sample size is consistent with non-probability sampling approaches commonly adopted in decision-support studies and exceeds the number of participants typically reported in FAHP-based prioritisation studies, which frequently involve between 10 and 20 experts. Table 3 summarises the demographic characteristics of the participants, indicating a predominance of professionals based in South America, a high level of educational attainment, a diverse range of job positions, and a balanced distribution of professional experience in construction projects. These characteristics may influence the depth of professional knowledge and perspective reflected in the findings, and also could potentially enrich the diversity of perspectives. This study followed established criteria to ensure credibility and transferability. Credibility was supported by grounding the questionnaire design in prior academic literature and established FAHP applications, while transferability was addressed by providing a clear description of the research context, methodology, and data collection process. Reliability was assessed using the consistency ratio (CR), with all matrices meeting accepted thresholds. Sensitivity analysis A sensitivity analysis was conducted adopting a criteria weight-perturbation method to assess the robustness and reliability of the FAHP-based prioritisation results. The analysis examined variations in the relative importance of the four main challenge categories (domain-specific, technological, adoption-related, and ethical). To this end, a set of scenarios was simulated to represent an equivalent condition to a pairwise comparison matrix at the criteria level, reflecting the relative weights among the challenge categories. Following the procedures outlined by Triantaphyllou and Sanchez (1997) and Wu et al. (2008), the importance of each challenge category was systematically increased and decreased by ±10%, ±20%, and ±30%. As an increase in the importance of one challenge category necessarily implies a proportional reduction in the importance of the remaining categories, the weights of the sub-challenges belonging to the non-modified domains were adjusted through normalisation. Accordingly, new weights were assigned to the sub-challenges based on their respective domains, and a global ranking was recalculated for each scenario. Results. This section presents the prioritisation of generative AI adoption challenges derived from the FAHP analysis. It also examines the robustness of the results through a sensitivity analysis. FAHP prioritisation within challenge categories Tables 4–7 summarise the aggregated triangular fuzzy number (TFN) matrices for each challenge category, together with the corresponding defuzzified weights (Mi), normalised weights (Ni), and consistency ratios (CR). The aggregation of individual judgements was performed using the geometric mean method. All aggregated matrices exhibited CR values below the accepted threshold of 0.1, indicating a satisfactory level of consistency in the collective expert judgements. Within the Domain-specific category, D2 emerged as the most critical sub-challenge (Ni 1⁄4 0.29), while D4 was ranked lowest (Ni 1⁄4 0.21). The range of normalised weights within this category was relatively narrow (0.29–0.21), indicating a comparatively balanced perception of criticality among its sub-challenges. In the Technological category, T2 obtained the highest normalised weight (Ni 1⁄4 0.25), whereas T5) was assigned the lowest weight (Ni 1⁄4 0.13). This category exhibited one of the widest spreads in normalised weights (0.25–0.13), suggesting greater variability in perceived criticality across technological sub-challenges. For the Adoption category, A1 ranked highest (Ni 1⁄4 0.28), while A3 ranked lowest (Ni 1⁄4 0.23). This category showed the narrowest range of normalised weights (0.28–0.23), indicating a relatively uniform perception of importance among adoption-related sub-challenges. In the Ethical category, E1 was identified as the most critical sub-challenge (Ni 1⁄4 0.34), while E3 received the lowest weight (Ni 1⁄4 0.21). This category displayed the widest range of normalised weights (0.34–0.21), indicating substantial variability in the perceived criticality of ethical sub-challenges. Overall ranking of generative AI adoption sub-challenges Table 8 presents the overall ranking of all sub-challenges across the four domains based on their normalised weights (Ni). As all sub-challenges were evaluated using the same criterion, their perceived impact on the adoption of generative AI tools, the resulting weights are directly comparable. The top three positions in the global ranking were occupied by E1, D2, and D1, belonging to the Ethical and Domain-specific categories. In contrast, the lowest-ranked sub-challenges were T3, T4, and T5, all from the Technological category. Adoption-related sub-challenges were predominantly positioned in the upper-middle range of the global ranking, while the Domain-specific category contributed three of its four sub-challenges to the top eight positions. Overall, the normalised weights ranged from 0.34 to 0.13, resulting in a total spread of 0.21 across all sub-challenges. Sensitivity analysis The resulting changes in the global ranking provided insights into how variations in the input weights influenced the prioritisation ranking of generative AI adoption sub-challenges. Across the four domains, sub-challenge E1 consistently remained in the first position in almost all scenarios, indicating a high level of robustness. Similarly, sub-challenges D1 and D2 demonstrated strong stability, remaining among the top-ranked positions under all perturbation scenarios. Detailed ranking variations across all scenarios are reported in Appendix B. Sub-challenges that maintained their positions or exhibited only minor variations across all scenarios were considered robust elements, indicating that their weights are relatively insensitive to changes in the criteria. By comparing the four matrices, it was possible to distinguish sub-challenges that were highly sensitive to weight perturbations from those that contributed most strongly to the robustness of the model. Overall, the sensitivity analysis confirms the robustness of the FAHP-based prioritisation, as key sub-challenges maintained stable ranking positions even under substantial criteria weight perturbations. Discussion. The FAHP results (Table 8) reveal a shift in technology adoption: practitioners perceive the primary challenges to generative AI not as technical limitations, but as systemic risks. The top ranking of data privacy, security, and legal liability suggests that the "black box" nature of generative AI creates a trust deficit that outweighs its functional benefits. Practitioners are aware of the potential risks associated with generative AI adoption, including data misuse and unauthorised access to sensitive project information. This is largely consistent with previous studies that have highlighted the significant legal and financial risks associated with AI deployment in construction and related industries. This indicates that the bottleneck for achieving the productivity growth target is a lack of legal and ethical frameworks. There is an urgent need for establishing robust industry-wide standards for data governance. The prominence of domain-specific and adoption-related sub-challenges shows the misalignment between general-purpose generative AI tools and the specialized requirements of the construction projects. This may be attributed to the persistent difficulties in acquiring professionals with the required skills and experience to effectively implement generative AI solutions. Likewise, the prominence of domain-specific and adoption-related challenges aligns with findings reported by Li et al. (2024) and Samsami (2024), who emphasised that many AI tools used in construction are trained on general-domain data, leading to limited contextual understanding and the need for frequent fine-tuning. In addition, the importance of adoption-related challenges is consistent with the work of Heo and Na (2025) and Smetana et al. (2024), who highlighted the lack of practical experience and structured educational pathways for implementing generative AI in construction projects. Interestingly, technologies challenges were ranked as the least critical factors. Participants perceived these issues as technical limitations that can be managed during implementation and that primarily affect task-level outputs rather than strategic decision-making. This is supported by the observations of Uddin et al. (2023) and Regona et al. (2022), who suggested that, despite existing technical limitations, construction organisations continue to adopt generative AI tools, particularly large language models, as these challenges can be addressed during implementation. Conclusion. This study identified and prioritised the key challenges associated with the adoption of generative AI in construction projects. The findings indicate that data privacy, security and legal liability; poor quality and availability of training data; and limitations related to construction-specific knowledge and language are the most critical challenges across the four domains. These results suggest that industry practitioners should prioritise the development of robust ethical and governance frameworks, as well as investing in high-quality domain-specific datasets and workforce upskilling initiatives. In particular, organisations should focus on aligning generative AI tools with construction-specific contexts, rather than relying on general-purpose models. Overall, the findings suggest that the successful adoption of generative AI requires a holistic and integrated approach that considers technical, organisational, contextual, and ethical challenges, as their relative importance may vary depending on the implementation context. This study provides a more nuanced understanding of generative AI adoption in construction projects by offering an empirical, expert-based prioritisation of adoption challenges through a combined SLR, FAHP, and sensitivity analysis approach. The findings deliver a robust and transparent ranking of these challenges, grounded in expert judgement and validated across multiple perturbation scenarios. While existing studies have explored generative AI opportunities and challenges in construction, much of this work remains conceptual or examines challenges in isolation, and often lacks empirical validation in real project contexts. To address these gaps, this study contributes to the body of knowledge by providing an updated and empirically grounded understanding of the challenges associated with the adoption of generative AI in construction projects. It advances existing research through the prioritisation of critical challenges based on predefined criteria and practitioners’ experiential judgements, as well as by assessing the robustness of the resulting rankings. From a practical perspective, the findings offer a transparent and actionable challenge prioritisation ranking that indicates the relative criticality and sequence in which challenges should be addressed, supporting construction managers and organisations in focusing their efforts during implementation. Methodologically, this research applies an integrated SLR–FAHP–sensitivity analysis approach to structure and prioritise adoption challenges in the emerging context of generative AI in construction, where such structured approaches remain limited. This study presents several limitations that also indicate directions for future research. The FAHP-based prioritisation relies on purposive expert judgements collected within specific professional contexts; therefore, the findings should be interpreted as decision-support insights rather than generalisable results. Future studies could expand the number of experts, the diversity of their profiles, and the geographical contexts considered, in order to enhance transferability. In addition, the prioritisation reflects expert perceptions at a specific point in time, while generative AI remains an emerging and rapidly evolving technology. Longitudinal research could examine how challenge prioritisation changes as industry experience and technological maturity increase. Finally, while this study focuses on prioritising adoption challenges, future research could extend the proposed SLR–FAHP approach by integrating complementary methods to explore prescriptive implementation strategies and managerial response pathways. CRediT: Alejandro Vidal Fernandez: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Validation, Visualization, Writing – original draft; Yixue Shen: Conceptualization, Methodology, Supervision, Validation, Writing – review & editing. Disclosure of interest The authors report there are no competing interests to declare. Author contributions Ethics and consent statement This study involved a non-interventional, anonymous questionnaire survey. Ethical approval was not required. Participation was voluntary, informed consent was obtained, and no identifiable personal data were collected, in accordance with the Declaration of Helsinki. Data availability statement The data that support the findings of this study are available from the corresponding author, Alejandro Vidal Fernandez, upon reasonable request. The data are not publicly available due to ethical considerations and the need to protect the privacy and confidentiality of the expert participants.