Aims: Accurate cost estimation is one of the fundamental prerequisites for successful planning and management of construction projects. The emergence of generative artificial intelligence (AI) models and online cost estimation systems has heightened the need to evaluate the performance of these tools in comparison with human expert judgment. Despite the proliferation of these technologies, no comprehensive study has simultaneously compared human experts, online systems, and generative AI models based on actual project costs. Existing research has largely focused on individual approaches in isolation, failing to provide an integrated comparative framework to guide practitioners. Moreover, the repeatability of AI-generated responses across multiple iterations with a fixed prompt has received limited attention, and the attitudes and trust levels of practicing engineers toward these novel tools remain underexplored. This research aims to address these gaps by systematically comparing the three aforementioned approaches in the context of the "Atashgah" building project in Semnan, Iran. Specifically, the study pursues four main objectives: (1) to evaluate differences in mean cost estimates among human experts, online systems, and generative AI models; (2) to assess the accuracy of each approach using MAE and MAPE relative to actual project cost, and measure their stability through coefficient of variation (CV); (3) to examine the repeatability of five AI models across 10 iterations with a fixed prompt; and (4) to investigate engineers' attitudes toward these tools and determine their level of trust in the outputs of online systems and generative AI.
Materials & Methods: This applied descriptive–analytical study adopted a survey-based comparative case study design. The case project was the Atashgah building, a six-story mixed-use (residential–commercial) structure located in Semnan, Iran, with a total floor area of 2,622 m². The building consists of a reinforced concrete frame with block-joist slabs and includes first-grade Iranian construction materials, a brick-and-stone façade, complete mechanical and electrical installations, a six-person elevator, and double-glazed windows. The actual construction cost (81.5 billion IRR), equivalent to 31.08 million IRR/m², served as the reference value for evaluating estimation accuracy. Data were collected through three parallel sources. First, a questionnaire was completed by 29 construction professionals representing different educational levels (65.51% bachelor's, 24.14% master's, 6.90% diploma, and 3.45% doctoral degrees), work experience (17.24% <5 years, 31.03% 5–10 years, 27.59% 11–20 years, and 24.14% >20 years), and professional roles (contractors, supervisors, site managers, and project managers). Participants estimated the project cost per square meter and reported their confidence in their own estimates (1–10), trust in online estimation systems (1–10), trust in AI-based estimations (1–10), trust in the official engineering tariff system (1–10), previous experience using online platforms and AI, and reasons for distrust through open-ended questions. The questionnaire demonstrated satisfactory validity (CVR = 0.82, CVI = 0.86) and reliability (Cronbach's α = 0.84). Second, eight active Iranian online cost estimation platforms (namaplan.com, dezharco.com, bkesh.ir, ahan100.com, karkoshte.com, pardissazeh.com, peymanu.com, and khanehtarh.ir) were evaluated. The same project specifications were entered into each platform according to its input requirements, and the estimated cost per square meter and total project cost were recorded. Because peymanu.com and khanehtarh.ir generated estimates based on a standard area of 100 m² using 2024 prices, their outputs were scaled to the project's actual area (2,622 m²) and updated to 2025 values using the official annual construction inflation rate of 47%. Third, five widely used generative AI models—ChatGPT, Gemini, Claude, DeepSeek, and Grok—were evaluated. These models were selected to represent different AI architectures (proprietary and open-source), geographic origins (U.S. and Chinese models), accessibility, and frequency of use reported by the surveyed engineers. All models received an identical prompt describing the complete project specifications, including building type, location, structural system, floor area, material quality, façade, elevator, building services, and window specifications. Each model generated 10 independent estimates using a new conversation for every iteration, resulting in 50 AI-generated responses. Web search functionality was disabled, and all evaluations were conducted using free-tier accounts. Estimation performance was assessed using the mean, standard deviation (SD), coefficient of variation (CV = SD/Mean × 100), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) relative to the actual project cost. Differences among estimation methods were examined using the Kruskal–Wallis test followed by Mann–Whitney U tests with Bonferroni correction. Engineers' confidence and trust ratings were summarized using descriptive statistics, while qualitative content analysis of the open-ended responses was conducted to evaluate the extent to which AI models considered local construction conditions, material quality, and market-related risks.
Findings: The results revealed that none of the three approaches individually possessed sufficient reliability for standalone application in real projects. Human expert estimates showed a mean of 30.83 million IRR/m², remarkably close to the actual cost of 31.08 million IRR/m² (MAPE = 0.80%), but exhibited considerable dispersion (CV = 34.64%), with estimates ranging from 10 to 65 million IRR/m² and a median of 30 million IRR/m². Despite this variability, engineers reported relatively high confidence in their own estimates (mean = 7.04 out of 10). Online systems exhibited the lowest reliability, with estimates ranging from 25 to 53.49 million IRR/m²—a 157% difference between the lowest and highest estimates. The mean estimate was 37.40 million IRR/m², with a standard deviation of 9.09 million and CV = 24.29%. This represented a 20.34% error relative to actual project cost, indicating a systematic tendency toward overestimation. Among the AI models, Grok demonstrated the best performance, with a mean estimate of 31.80 million IRR/m² (MAPE = 2.32%, CV = 5.32%), and a narrow range of 71–99 billion IRR total cost. Gemini ranked second with MAPE = 12.48% and CV = 19.4%. ChatGPT and DeepSeek showed moderate performance with MAPEs of 21.66% and 22.72%, and CVs of 14.09% and 34.14%, respectively. Claude performed poorly, with MAPE = 63.77%, CV = 71.23%, and high instability across iterations (range: 9.5–75 billion IRR); it also refused to provide numerical estimates in 3 out of 10 iterations. Engineers expressed low trust in non-human approaches, with mean trust scores of 4.89 for the official tariff system, 4.68 for online sites, and 4.67 for AI—all below 5 out of 10. Only 35.71% had used AI before, and merely 20.69% had used online sites. The Kruskal–Wallis test confirmed significant differences among the three approaches (H = 24.6, df = 2, p < 0.001). However, post-hoc Mann–Whitney tests with Bonferroni correction revealed that only the difference between online systems and AI models was statistically significant (p < 0.001). Qualitative analysis of AI responses indicated that Gemini and DeepSeek showed greater attention to local conditions (e.g., "price fluctuations in Iran" and "cost differences in provincial cities"), while ChatGPT and Grok provided standard, repetitive explanations, and Claude offered general, non-specialized content.
Conclusion: This study concludes that no single approach to construction cost estimation is sufficiently reliable for independent application. Human experts, while confident, exhibit significant variability. Online systems lack both accuracy and consistency. AI models show promise but vary widely in performance and stability. The hybrid "human-in-the-loop" approach is proposed as the most effective strategy, wherein a stable AI model (such as Grok) generates an initial estimate, which is then validated, refined, and completed by human experts based on their contextual knowledge and experience. Engineers and project managers are advised to evaluate AI model stability across multiple iterations and select models with CV < 15% for practical applications. Professional engineering organizations should develop formal guidelines and training programs to enhance trust and adoption of AI technologies. Future research should expand the sample size, replicate the study across diverse projects and cities, and develop explainable AI tools to improve transparency and trust. Despite rapid AI advancements, engineering judgment remains indispensable, and new technologies should serve as complementary tools augmenting, not replacing, human expertise.