AI-Driven Automation of Construction Cost Estimation: Integrating BIM with Large Language Models

Abstract

The construction industry faces challenges in estimating costs because the processes are time-consuming and involve a high likelihood of making errors. For instance, quantity take-offs are often inaccurate, and there is not a simple way to integrate data from Building Information Modeling (BIM) platforms and cost databases. This study introduces a framework that utilizes the Model Context Protocol (MCP) to ensure seamless integration between large language models (LLMs) and BIM models through Autodesk Revit in order to enable fully automated cost estimation workflows. The developed system combines an AI-powered MCP server with cost databases that are standard in the industry, such as the 2025 Craftsman National Building Cost Manual and the ZIP code-based location modifiers. This system enables LLMs to automatically obtain quantities from BIM models, match components to cost items, make regional changes, and make professional cost estimates. A case study of estimating the cost of an electrical system shows that the framework can reduce estimation time from 2.5–3.5 h (manual baseline) to 42.3 ± 3.7 s (n = 5 runs, warm start), representing a 98.6% efficiency gain, while being more accurate with respect to industry standards. The system processed 187 BIM elements in three component groups (receptacles, conduits, and panels). It automatically matched them to the right cost database items, used location-specific modifiers for ZIP code 01003, and made a full cost estimate of USD 13,945.81 with detailed breakdowns and a percent difference of %5.1 of the manual estimation. This research enhances automation in construction by developing a methodology for AI-BIM integration using standardized protocols, shows the practical application of AI in construction workflows, and provides empirical evidence of the advantages of automation in cost estimation processes. The results indicate that MCP-based AI integration presents a novel approach for construction automation, delivering improvements while applying professional standards of accuracy and availability.

1. Introduction

The construction industry has historically lacked the adoption of digital technologies and productivity gains compared to other industries, such as automobile manufacture and production [1]. Construction workflows are dynamic, labor-intensive, and easily involve errors, especially when it comes to the cost estimation procedures that serve as the cornerstone of project financial planning and control [2]. Construction project cost estimation needs manual labor, including quantity take-off from drawings, price lookup in cost databases, location adjustments, and compilation of thorough estimates [3]. These procedures take a lot of time and money and are still subject to human error [3].

In the architecture, engineering, and construction (AEC) sectors, Building Information Modelling (BIM) has become a game-changing technology that makes it possible to digitally represent the functional and physical aspects of buildings [4,5]. By integrating 3D models with a wealth of project data, BIM enables automated quantity take-off, providing advantages over manual measurement [6]. However, there are still a lot of gaps in the way BIM data can be seamlessly connected to downstream cost estimation processes, especially when it comes to automating component-to-cost matching, applying location modifiers, and producing expert deliverables [6].

Large language models (LLMs) have shown impressive capabilities in natural language processing, reasoning, and task automation in recent advances in artificial intelligence (AI) [7]. With little assistance from humans, these AI systems can process both structured and unstructured data, comprehend complicated instructions, and carry out multi-step workflows. Nonetheless, the absence of standardized interfaces and protocols for contextual data exchange has hindered the successful integration of AI capabilities with domain-specific tools such as BIM software [8].

By facilitating smooth bidirectional communication between AI applications and external data sources, the Model Context Protocol (MCP), which was introduced as a standardized communication framework, solves these integration issues [8]. The MCP offers structured primitives for tools, resources, and prompts, enabling AI systems to dynamically access data, find capabilities, and invoke functions across various platforms [8]. This protocol transforms the traditional “M × N integration problem” into a more manageable “M + N problem,” significantly reducing integration complexity [9]. Claude Desktop’s User Interface Layer lets practitioners start workflows using natural language commands without writing. Claude, a big language model, manages requests, performs multi-step plans, and creates structured reports in the AI Processing Layer [9]. The Integration Protocol Layer implements the MCP, which establishes a standard communication bridge between AI and external tools [10].

Cost estimation in the construction industry is challenging due to various issues that make workflows complex, dynamic, and challenging to repeat reliably [11]. Manual process inefficiencies, such as quantity take-offs from drawings, cost database lookups, and the manual application of location-based cost factors, predominate in current practices [11]. Data fragmentation makes this issue worse by requiring estimators to manually fill in the gaps left by disconnecting rich data from cost databases, regional modifiers, and BIM models [12]. Additionally, current solutions rely on proprietary plugins or custom scripts, which are challenging to scale and maintain due to significant integration barriers [13]. As a result, the crucial following processes of component classification, cost–item matching, and report generation are still mostly manual or semi-automated, even when BIM is utilized for automated quantity extraction [6]. These difficulties show how urgently the sector needs creative solutions that make use of cutting-edge technologies to change cost estimation from a disjointed, manual process to an automated, intelligent, and integrated one.

This research aims to develop an empirical AI-driven framework that addresses the integration challenges in construction cost estimation. To achieve this aim, the study is guided by the following research questions:

RQ1: 

How can a standardized protocol MCP effectively bridge the interoperability gap between BIM platforms, AI systems, and cost databases to enable automated data exchange for cost estimation?

RQ2: 

What level of accuracy and reliability can LLM-orchestrated cost estimation achieve compared to manual professional standards?

RQ3: 

What efficiency gains (time reduction) can be achieved through AI-driven automation of the end-to-end cost estimation workflow?

To address these research questions, the study pursues four specific objectives:

• Framework Development: Design a scalable four-layer architecture using the MCP to enable LLMs to communicate directly with BIM platforms such as Autodesk Revit and industry cost databases.
• Practical Demonstration: Implement and demonstrate an end-to-end electrical cost estimation case study to prove the practical feasibility of the proposed framework.
• Performance Validation: Quantitatively validate the system’s performance by measuring accuracy (cost variance from manual estimates) and efficiency (time reduction) against manual professional processes, establishing precise performance benchmarks.
• Methodology Documentation: Document the entire workflow as a reproducible methodology, providing the industry with a workable adoption blueprint and concrete evidence for AI-BIM automation.

The remainder of this paper is organized as follows: Section 2 reviews the relevant literature on BIM-based cost estimation, AI in construction, and integration protocols. Section 3 details the research methodology and the proposed MCP-based system architecture. Section 4 presents a comprehensive case study on electrical system cost estimation. Section 5 analyzes the results and validates the system’s accuracy and performance. Section 5 discusses the implications, limitations, and future research directions. Finally, Section 6 concludes the paper by summarizing the key findings and contributions.

2. Literature Review

This review examines the existing body of knowledge across three critical domains: the capabilities and limitations of BIM in cost estimation, the emerging application of AI in construction, and the key role of integration protocols. By addressing these areas, the research gap addressed by this study is clearly identified.

2.1. BIM-Based Quantity Take-Off and Cost Estimation

BIM has reshaped quantity take-off (QTO) processes from the traditional manual quantification from 2D drawings, which is error-prone and time-consuming [6]. This transformation has been achieved by enabling the automated extraction of accurate quantities directly from information-rich 3D models, with tools such as Autodesk Revit’s Schedules being widely used for this purpose [6]. Research has extensively explored both the potential and pitfalls of BIM-based QTO [14]. Monteiro and Martins [14] established modeling guidelines to ensure that BIM models contain sufficient detail for accurate quantity extraction, highlighting that model quality is a prerequisite for automation. The integration of these quantities with cost estimation has been a valuable focus [14]. Lee et al. [15] proposed an ontology-based approach to bridge the semantic gaps between BIM model information and cost database requirements. Similarly, Ma and Liu [16] developed a semi-automatic system for generating specification-compliant cost estimates from Industry Foundation Classes (IFC) data. These studies demonstrate feasibility and reveal a common limitation, as manual intervention remains necessary for component classification, cost matching, and estimate compilation.

A persistent challenge is the incompatibility between BIM-based quantity definitions and the structure of cost databases [17]. For instance, a cost database might define grout quantity by wall area, while the BIM model calculates it by volume, necessitating a complex translation logic [18]. Furthermore, BIM models often fail to represent temporary works or construction methodologies, which must be manually added to estimates [18]. Recent studies have focused on improving the accuracy of QTO for compound elements and developing automated high-accuracy systems [6,17]. However, these efforts primarily target the initial quantity extraction phase, leaving the subsequent steps of the cost estimation workflow largely unautomated [19].

2.2. Artificial Intelligence in Construction Cost Estimation

The application of artificial intelligence in construction has expanded into areas such as project scheduling, safety monitoring, and cost prediction [20]. Traditional machine learning (ML) techniques, including Neural Networks, have shown promise in predicting construction costs based on historical project data [21]. However, these ML approaches typically require large, labeled datasets for training and may struggle with novel project types or rapidly changing market conditions [21].

A transformative development has been the advent of LLMs and generative AI. These models demonstrate exceptional capabilities in natural language understanding, reasoning, and code generation, allowing them to execute complex, multi-step tasks [22]. LLMs leverage vast pre-trained knowledge and can be adapted through conversational interfaces or few-shot learning, making them highly suitable for domains with limited standardized datasets, such as construction [23,24].

Despite this potential, the practical application of LLMs in construction remains in its early stage [23,24]. Most existing research presents theoretical frameworks or proof-of-concept demonstrations rather than production-ready implementations [23]. Gamil et al. [25] studied a critical barrier in the integration challenge, connecting the general-purpose capabilities of LLMs to the domain-specific tools, databases, and logic required in construction workflows. This gap motivates the investigation of standardized integration protocols to enable the robust and scalable deployment of LLMs in construction [25].

2.2.1. Limitations of Conventional AI Approaches

While traditional AI approaches have demonstrated value in construction cost estimation, each method presents inherent limitations that constrain its applicability for end-to-end workflow automation. Rule-based systems and expert systems, though transparent and interpretable, require extensive manual encoding of domain knowledge and struggle with the semantic variability in component naming conventions across different BIM models [15]. The rigidity of predefined rules limits adaptability when encountering unfamiliar components or unconventional project specifications.

Ontology-driven reasoning frameworks address some semantic interoperability challenges by formalizing relationships between BIM concepts and cost database structures [15]. However, as Yin et al. [26] demonstrated, ontology-based approaches require substantial upfront investment in ontology development and population and cannot efficiently handle the project-specific terminology and customized properties that vary from one project to another [26]. Their research found that existing BIM ontologies and terminology models fail to cover customized properties, requiring additional ontology population techniques to absorb project-specific concepts, a limitation that significantly constrains practical deployment.

Conventional machine learning models, including Support Vector Machines (SVMs), Decision Trees, and Neural Networks, have shown promise in cost prediction tasks. Elmousalami [27] conducted a comprehensive review comparing twenty machine learning models for construction cost estimation, finding that while hybrid models achieve Mean Absolute Percentage Error (MAPE) values between 7 and 15%, these approaches require large, labeled training datasets and are primarily designed for predictive tasks rather than workflow orchestration. Wang et al. [28] noted that SVMs, DTs, and RFs have inherent limitations, including overfitting for regression problems and an inability to handle the multi-step reasoning required for complete estimation workflows. Deep Neural Networks (DNNs) can achieve higher accuracy with sufficient data but perform poorly with small datasets typical of specialized construction domains [27], as the computational requirements and training time make them impractical for project-specific adaptation.

2.2.2. The Advantage of Large Language Models

LLMs present distinct advantages that address these limitations. First, LLMs possess semantic reasoning capabilities that enable flexible interpretation of heterogeneous naming conventions without requiring a rigid, predefined mapping logic [13]. This capability allows LLMs to interpret BIM component names like “3/4” EMT Conduit” and match them to cost database entries with varying nomenclature. Second, LLMs can perform few-shot learning and leverage pre-trained knowledge bases, eliminating the need for extensive project-specific training data that constrains conventional ML approaches [29]. Third, LLMs uniquely support conversational interfaces that enable practitioners to initiate complex workflows through natural language commands, democratizing access to automation capabilities without requiring programming expertise [26]. Fourth, the orchestration capability of LLMs allows them to dynamically coordinate multiple tools, databases, and calculation processes within a single coherent workflow, a capability that rule-based or single-purpose ML models cannot replicate [8].

These characteristics position LLMs not as replacements for specialized prediction algorithms, but as orchestration layers that can reason through the component-matching process, coordinate diverse data sources, and generate professional deliverables while maintaining interpretability through natural language explanations [8].

2.3. Model Context Protocol and Integration Frameworks

The MCP has been introduced as a standardized framework to address the fragmentation in the AI tool ecosystem [13]. The MCP provides a unified communication layer that enables AI applications to interact seamlessly with external resources, tools, and data sources [13]. Its architecture is built on three core primitives: prompts (reusable conversational templates), resources (data references), and tools (functions an AI can invoke) [13]. Figure 1 is an MCP architecture diagram showing client–server interaction across local and remote systems.

This standardization is essential; as shown in Figure 2, it transforms integration complexity from an “M × N” problem to an “M + N” problem [8]. In the context of construction, where multiple specialized software tools are used, this reduction in complexity offers substantial practical benefits. The MCP supports dynamic capability discovery, allowing AI agents to query servers at runtime to understand available tools without prior configuration, enabling adaptive and context-aware workflows [13].

Recent work by Addepalli et al. [30] emphasizes the MCP’s role as a “universal translator” within cloud ecosystems, enabling secure and scalable interaction between AI systems and diverse enterprise data sources. However, while the MCP has seen adoption in software development tools such as Cursor Integrated Development Environment (IDE), its application in the construction domain remains largely unexplored. Furthermore, Hou et al. [8] have analyzed security threats and future directions for the MCP, highlighting the need for robust authentication and data privacy, considerations that are paramount for handling sensitive construction and cost data.

2.4. Synthesis and Identified Research Gaps

A comprehensive literature review shows that BIM-based quantity take-off and AI in construction have advanced when they work in parallel rather than together [15]. Existing approaches, whether commercial plugins or research prototypes, use custom, brittle integrations that are hard to maintain and scale, or address only isolated workflow steps [31]. This fragmentation highlights multiple interconnected research gaps that define a compelling research agenda [31]. First, no previous effort has used a standardized framework such as the MCP to overcome the fundamental interoperability problem between BIM platforms, AI agents, and cost databases [31]. Second, this fragmentation also creates an end-to-end automation gap; fully autonomous workflows from model interrogation to professional report preparation are rare [31]. Third, big language models’ ability to recognize construction context, reason about component matching, and orchestrate multi-step estimate tasks is untested in real-world circumstances [16]. Fourth, many offered solutions are proprietary tools or proof-of-concept demonstrations without a transparent, flexible technique for wider adoption [16]. Finally, a lack of quantitative benchmarking of automated estimating accuracy against manual norms undermines practitioner confidence [16]. This research addresses these shortcomings by presenting and testing an MCP-based architecture that allows an LLM to autonomously perform an end-to-end cost estimating workflow, creating a reproducible and empirically supported paradigm for scale construction automation.

3. Methodology

This research employed the design science research (DSR) methodology [32]. DSR is a problem-solving paradigm that focuses on the creation and evaluation of innovative artifacts, such as models, methods, or frameworks, to address defined practical challenges [32]. The DSR approach was selected because it provides a rigorous framework for developing and evaluating innovative technical solutions in domain-specific contexts, aligning with similar methodological approaches in BIM and construction automation research [17,21].

The research design followed the DSR cycle, comprising four sequential phases, each mapped to specific research questions:

Phase 1: Requirements Analysis and Framework Design (RQ1)

This phase involved a systematic literature review to identify gaps in existing BIM–cost estimation integration approaches, followed by architectural design of the MCP-based framework. Requirements were derived from practitioner workflows and technical constraints of BIM software APIs.

Phase 2: System Implementation and Integration (RQ1)

Implementation involved developing custom MCP servers for Autodesk Revit connectivity, integrating the Craftsman 2025 cost database, and configuring the Claude AI orchestration layer. All system components were developed using documented APIs and standard protocols to ensure reproducibility.

Phase 3: Case Study Application (RQ2, RQ3)

A representative commercial building electrical system was selected as the validation testbed. Selection criteria included the following: (a) moderate complexity with multiple component categories, (b) the availability of a complete BIM model with an accurate geometry, and (c) the applicability of standard cost database items. The electrical system encompassed receptacles, conduit networks, and distribution panels, components commonly encountered in commercial construction.

Phase 4: Validation and Benchmarking (RQ2, RQ3)

Validation employed a multi-faceted methodology:

Quantity Extraction Accuracy: Automated counts and measurements were compared against a manual Revit model audit.

Cost Database Matching: An expert estimator validated the AI’s database item selection for specification and context appropriateness.

Calculation Integrity: Manual recalculation of sample line items verified adjustment and markup formula accuracy.

Overall Estimate Accuracy: The AI-generated total cost was compared to a manually prepared estimate using identical inputs, with variance benchmarked against industry-standard accuracy tolerances for detailed estimates (±5–10% at 90% confidence) as per established construction estimation practice (Sherif et al., 2011) [33].

The process was executed in four sequential phases: requirements analysis and framework design, system implementation and integration, case study application, and rigorous validation. To enable a quantitative investigation, a case study approach was adopted, focusing on the electrical system of a commercial building project. This scope allowed for an in-depth examination of the automation workflow of its performance against manual industry standards.

3.1. System Architecture

The developed system is built from a four-layer architecture (as shown in Figure 3) designed to ensure modularity and scalability.

(1)

The User Interface Layer provides access through the Claude Desktop application, allowing practitioners to initiate workflows via natural language commands without requiring programming skills.

(2)

The AI Processing Layer utilizes Claude 3.5 Sonnet (July 2025 release) as the central orchestration engine. The model configuration employs the following inference parameters:

• Temperature: 0.3 (efficient output for reproducible calculations);
• Top-p: 1.0 (no nucleus sampling restriction);
• Max tokens: 64,000 (sufficient for comprehensive report generation);
• Tool-calling: Native Claude tool use with structured JSON schema validation.

The system implements several guardrail constraints to ensure output reliability:

Output Schema Validation: All cost calculations are validated against a predefined JSON schema requiring numeric types for costs, quantities, and percentages.

Calculation Verification: The LLM performs internal consistency checks, verifying that extended costs equal unit costs multiplied by quantities (tolerance: ±USD 0.01).

Database Match Confirmation: Each component-to-cost-item match includes confidence indicators and specification alignment verification.

Error Handling: Unmatched components trigger explicit user notification rather than silent failure.

The orchestration prompts and system instructions are provided in Appendix A to enable replication.

(3)

The Integration Protocol Layer implements the MCP, creating a standardized communication bridge between the AI and external tools. A custom MCP server was developed for Autodesk Revit, exposing the tool endpoints shown in

Table 1. MCP Server Tools for BIM Model Interrogation.

Tool Name	Function	Input Parameters	Output
get_current_view_info	Retrieves active view metadata	None	View name, type, scale
get_current_view_elements	Extracts elements by category	modelCategoryList[]	Element array with properties
AI_element_filter	Advanced element filtering	filterCategory, includeInstances	Filtered element set
get_selected_elements	Retrieves user-selected elements	ElementsList[]	Selected element details

:

All tool responses conform to a standardized JSON schema, enabling consistent parsing regardless of element type. Error handling implements graceful degradation: connection failures trigger retries logic (3 attempts, exponential backoff), while query errors return structured error objects with diagnostic codes.

• Database Considerations:

The Craftsman National Building Cost Manual 2025 is a proprietary publication requiring licensed access. For scientific replication purposes, an anonymized subset of the electrical cost database structure (without actual pricing) is available in the Appendix A. In addition, researchers may substitute equivalent cost databases by matching the required schema: Category, Description, Material/Labor/Equipment Costs, Unit, and Crew Information

Security Implementation: The security protocol followed in this tool development focus on main three areas:

• Local-only MCP connections (no remote access);
• Read-only Revit model access (no write operations);
• No persistent storage of cost data outside session.

(4)

Finally, the Data Sources Layer encompasses the three core inputs: the Autodesk Revit BIM model containing the electrical design, the Craftsman National Building Cost Manual 2025 database with unit costs for over 13,000 electrical items, and the Craftsman ZIP Area Modifiers 2025 database for location-based cost adjustments. The Craftsman 2025 instructions provided industry-standard cost data, with project-specific parameters set to 01003 (Amherst, MA), 10% indirect cost, and 10% profit margin. These inputs created a realistic automated workflow testbed.

Table 2. Summary of case study data sources.

Data Source	Description	Key Metrics/Content
Revit Model	Commercial store electrical system	187 elements across receptacles, conduits, panels
Cost Database	Craftsman National Building Cost Manual 2025 (Electrical)	13,395 items; Material, Labor, Equipment Costs
Location Modifier DB	Craftsman ZIP Area Modifiers 2025	40,777 ZIP codes; Material, Labor, Equipment Factors
Project Parameters	User-defined inputs	ZIP: 01003; Indirect: 15%; Profit: 10%

provides a summarization of the data sources in the case study.

3.2. Automated Workflow Process

The automated workflow represents the core operational logic of the framework, wherein the LLM orchestrates a complete, end-to-end estimation pipeline through structured interaction with external tools via the MCP. This process is initiated by a single natural language user instruction and proceeds autonomously through five integrated phases, transforming raw BIM data into a validated cost estimate and professional report. Figure 4 shows the phases of the set-up of the project instructions using Claude AI.

3.2.1. Phase 1: Model Interrogation and Quantity Extraction

The LLM initiates the workflow by calling the MCP server connected to the active Autodesk Revit session. It first retrieves context about the active view and then executes a filtered query to extract all model elements belonging to relevant electrical categories (e.g., Electrical Fixtures, Conduits, Electrical Equipment). For each element, the system parses key properties, including family, type, and geometric data. Quantities are computed through a category-specific extraction logic, with all measurements in Imperial units (feet) to align with US cost database conventions.

Discrete Element Counting (Receptacles, Panels):

Elements are counted by aggregating unique Element IDs within each category. To prevent misclassification between visually similar elements (e.g., junction boxes vs. receptacles), the system applies a disambiguation rule hierarchy:

• Family Name Match: Elements must belong to recognized families (e.g., “Duplex Receptacle” family).
• Type Parameter Verification: Amperage and configuration parameters must match expected values (e.g., “20A” and “Duplex”).
• Category Exclusion: Elements in ambiguous categories (e.g., “Electrical Fixtures” containing both receptacles and junction boxes) are filtered by family name before counting.

Linear Element Measurement (Conduit):

Conduit length is calculated using the following protocol:

Segment Extraction: Each conduit segment is retrieved from the element length parameter value.

Run Aggregation: All segments within the same conduit system are summed.

Unit Conversion: Total length in feet is converted to CLF (cost per hundred linear feet) by dividing by 100.

Handling of Complex Geometries:

Fittings: Fittings (elbows, connectors, and junction boxes) are counted separately as discrete items when modeled; unmodeled fittings are not captured (noted as limitation).

Junction Losses: The system does not apply deductions at junctions; total run length represents gross conduit material, consistent with typical estimating practice, where waste factors are applied separately.

Validation Against Native Revit Schedules:

As a quality assurance measure, automated quantities were cross-checked against Revit’s native Schedule/Quantities feature:

• Receptacle count: Automated = 30, Revit Schedule = 29;
• Conduit length: Automated = 2450 LF, Revit Schedule = 2526.5 LF (discrepancy: 3.0%, attributed to segment aggregation differences);
• Panel count: Automated = 2, Revit Schedule = 2 (exact match).

These cross-checks informed the accuracy assessment reported in Section 5.1.

3.2.2. Phase 2: Database Integration and User Parameterization

Concurrently, the system loads and indexes the structured external databases. The Craftsman cost manual is parsed to enable efficient semantic search, while the location modifier database is prepared for instant lookup. The LLM then prompts the user for the essential project parameters required for final calculations: the project ZIP code (to retrieve location factors) and the desired percentages for indirect costs and profit margin. This phase ensures that the estimate is grounded in both standard cost data and project-specific adjustments.

3.2.3. Phase 3: Intelligent Component–Cost Matching

This phase applies the reasoning capability of the LLM to map extracted BIM components to appropriate line items in the cost database via a formalized, multi-stage matching algorithm. For each component type, the AI executes the following structured procedure:

• Candidate Retrieval: A keyword-based search identifies potential cost items from the relevant database category.
• Ranking/Scoring: Each candidate is scored using a composite metric:
  $$\mathrm{S}={\mathrm{w}}_{\mathrm{s}}\cdot \mathrm{sim}({\mathrm{D}}_{\mathrm{c}},{\mathrm{D}}_{\mathrm{d}}\mathrm{b})+{\mathrm{w}}_{\mathrm{p}}\cdot \mathrm{match}({\mathrm{P}}_{\mathrm{c}},{\mathrm{P}}_{\mathrm{d}}\mathrm{b})$$

  (1)

  where $\mathrm{sim}$ is the semantic similarity (e.g., cosine similarity of embedding vectors) between the component description, ${\mathrm{D}}_{\mathrm{c}}$, and the database entry, ${\mathrm{D}}_{\mathrm{d}}\mathrm{b}$; $\mathrm{match}$ is the proportion of exact specification matches (e.g., size, amperage, and voltage); and weights ${\mathrm{w}}_{\mathrm{s}}=0.7$ and ${\mathrm{w}}_{\mathrm{p}}=0.3$ prioritize semantic understanding while ensuring technical accuracy. Candidates scoring below a threshold $\mathrm{S}<0.8$ are discarded.
• Tie-Breaking and Conflict Resolution: If multiple candidates exceed the threshold, the item with the highest $\mathrm{match}$ sub-score is selected. If scores remain tied, the item with the lower unit cost (yielding a conservative estimate) is chosen.
• Controlled Fallback: If no candidate meets the threshold, e.g., due to missing specifications in the BIM model or the absence of a compatible database entry, the component is flagged for manual review. A generic placeholder item (e.g., “Electrical Component, Unspecified”) is assigned with a unit cost marked as “TBD,” allowing the workflow to proceed while ensuring transparency.

The selected cost item provides the baseline material, labor, and equipment unit costs, completing the critical link between design information and financial data.

3.2.4. Phase 4: Cost Calculation and Synthesis

With matched items and user parameters defined, the system executes a deterministic calculation engine. Unit costs are first adjusted by their respective location-based modifiers (Material, Labor, and Equipment). Extended costs are computed for each component (Adjusted Unit Cost × Quantity). These are aggregated into material, labor, and equipment subtotals. Markups are then applied in the standard industrial sequence: indirect costs are added to the direct cost total, profit is calculated on the resulting sum, and sales tax is applied where applicable to yield a final grand total.

3.2.5. Phase 5: Professional Reporting and Visualization

The final phase translates the computational results into a stakeholder-ready deliverable. The LLM generates a comprehensive, interactive HTML dashboard. This report features an executive summary highlighting the total cost, a detailed line-item table with full quantity and cost transparency, visual charts breaking down costs by system and type, and a clear summary of all applied assumptions and modifiers. The output is designed for both on-screen review and professional-quality PDF export, completing the transformation from model data to a decision-support document.

This seamless, LLM-orchestrated workflow demonstrates how the MCP enables the integration of discrete tools, BIM software, cost databases, and calculation logic into a single, automated sequence, bridging the gap between design information and financial estimation.

A multi-faceted validation methodology was used to ensure system output confidence. Quantity Extract: Automatic counts and length calculations were compared to a rigorous manual Revit model audit to determine accuracy. Item Cost Matching: An expert estimator validated the AI’s database item selection for each component for specification and context. Manually recalculating a sample of line items verified the adjustment and markup formulas’ Calculation Integrity. Finally, Overall Estimate Accuracy was measured by comparing the AI-generated total cost to a humanly prepared estimate using the same confirmed inputs and comparing the variation to industry-standard accuracy tolerances for detailed estimates. The process and output met professional standards with this extensive validation technique.

3.3. Performance Measurement Protocol

To ensure a rigorous comparison between automated and manual workflows, a standardized measurement protocol was established.

Automated System Time Measurement:

• Hardware Configuration: Intel Core i7-12700K (12 cores, 3.6 GHz base), 32GB DDR5 RAM, NVMe SSD storage, Windows 11 Pro.
• Software Environment: Autodesk Revit 2024, Claude Desktop Application (Claude 3.5 Sonnet, October 2024 release).
• Measurement Protocol: Five consecutive runs were executed after initial warm-up (database pre-loaded into memory). Time was measured from user command initiation to complete HTML report generation, including (a) MCP server initialization, (b) Revit model interrogation, (c) cost database parsing, (d) component matching, (e) calculations, and (f) report generation.
• Results: Mean execution time: 42.3 s (SD = 3.7 s, n = 5), range: 38.1–47.2 s.

Manual Baseline Construction:

• Estimator Profile: Professional cost estimator with 8+ years of commercial electrical estimation experience.
• Method: Standard industry practice using Revit Schedules for QTO, manual lookup in Craftsman 2025 database, spreadsheet-based calculations with location modifiers, and formatted report preparation.
• Inputs/Assumptions: Identical to automated system (same BIM model, cost database edition, ZIP code 01003, 15% indirect costs, 10% profit margin, 0% sales tax).
• Timing: Three independent manual estimates were completed, with times recorded: 2.5, 3.0, and 3.5 h (mean: 3.0 h).

4. Case Study

To empirically validate the proposed framework, a comprehensive case study was conducted focusing on the electrical system of a representative commercial retail building. This project was selected for its inclusion of standard MEP components and moderate complexity, providing a realistic testbed for evaluating the automated workflow from model interrogation to final cost reporting. The following sections detail the project context, the stepwise execution of the automated process, and the quantitative results obtained.

4.1. Case Study Context and Model Description

The case study involved the construction of an Autodesk Revit model of a single-story commercial store building that integrates electrical systems. The project location was set to be ZIP code 01003 (Amherst, Massachusetts), requiring regional cost adjustments. For this study, attention was directed to the electrical system, which included three primary component categories: specification-grade duplex receptacles, Electrical Metallic Tubing (EMT) conduit runs, and electrical distribution panels. As illustrated in the 3D view of the Revit model (Figure 5), these components were fully modeled with accurate geometry, spatial coordinates, and type properties, providing a rich data source for automated quantity take-off. The Revit model contains comprehensive electrical systems, including 29 electrical receptacles (duplex 15 amp outlets), 156 conduits totaling 2526.5 linear feet of 3/4” EMT conduit, and 2 electrical distribution panels (125 amp, 20-pole capacity). User-specified parameters include 10% indirect costs for overhead and project management, 10% profit margin, and 0% sales tax (assumed to be incorporated in other rates or exempt).

4.2. Execution of the Automated Workflow

The workflow was initiated by a single natural language command issued to the Claude AI agent within the configured desktop application. The agent has a predefined role and systematic instructions for cost estimation, which outline the phased procedure it follows.

4.2.1. Phase 1: Quantity Extraction

The AI system initiates the workflow by connecting to the active Revit model through the MCP. The get_current_view_info tool retrieves the view context: View Name: “{3D}”, confirming that the model contains a three-dimensional geometry suitable for quantity take-off. Element extraction utilizes the get_current_view_elements tool with specified filters for electrical categories. The system processes returned JSON data containing 187 total electrical elements. Parsing and categorization organizes elements by type: 30 elements in the “Electrical Fixtures” category (receptacles), 155 elements in the “Conduits” category (conduit runs), and 2 elements in the “Electrical Equipment” category (panels). Conduit length calculation demonstrates the geometric processing capabilities. Each conduit element contains Start and End properties with X, Y, Z coordinates. For example, for one conduit segment: Element ID 868519, Start: (135.0, −6.9, 2.0), End: (135.0, 0.8, 2.0). Calculated length: $\sqrt{\left[\right(135.0-135.0{)}^2 + (-6.9-0.8{)}^2 + (2.0-2.0{)}^2]}= 7.75$ feet. Aggregating all 155 segments yields the total conduit length: 2450 linear feet. Unit conversion to CLF (cost per hundred linear feet): 2450/100 = 24.50 CLF, as cost databases typically price conduits per CLF.

4.2.2. Phase 2: Database Integration

Cost database loading utilizes a Sheets library to parse the Excel workbook. The Craftsman Electrical 2025.xlsx file contains 13,395 cost items organized into columns: Category, Description, Material Cost per Unit, Labor Cost per Unit, Lump Sum Cost, Manhours per Unit, Equipment Cost per Unit, Unit of Measure, Crew Code and Detail, and Crew Code. Database indexing creates efficient search structures. A category index maps keywords to arrays of matching items, enabling rapid filtering. For example, the keyword “receptacles” maps to 247 database entries for various receptacle types. Description indexing further organizes items within categories for refined searching. As shown in

Table 3. Pseudocode algorithms for MCP-based automated cost estimation.

Step	Pseudocode	Description
Database Integration and Modifier Retrieval Algorithm
1	LOAD electrical_db FROM “Electrical 2025.xlsx”	Initialize connection to cost database
2	PARSE electrical_db.SHEET[“CostbookExport”] TO json_data	Extract structured cost data
3	LOAD zip_db FROM “Craftsman Zip Area Modifiers 2025.xlsx”	Initialize connection to location database
4	PARSE zip_db.SHEET[“2025CraftsmanZipBasedAreaM”] TO zip_data	Extract location adjustment factors
5	FIND zip_record IN zip_data WHERE zip_record[“ZIP_CODE”] = project_zip	Locate project-specific modifiers
6	mat_multiplier = 1 + (zip_record[“Material_factor”]/100)	Calculate material adjustment
7	lab_multiplier = 1 + (zip_record[“Labor_factor”]/100)	Calculate labor adjustment
8	eqp_multiplier = 1 + (zip_record[“Equipment_factor”]/100)	Calculate equipment adjustment
9	RETURN json_data, mat_multiplier, lab_multiplier, eqp_multiplier	Output data and multipliers
Component–Cost Matching Algorithm
1	FOR EACH component IN extracted_bim_elements	Iterate through all BIM components
2	category = IDENTIFY_CATEGORY(component)	Determine component classification
3	specifications = EXTRACT_SPECS(component)	Extract technical specifications
4	keyword_list = GENERATE_KEYWORDS(category, specifications)	Create search keywords
5	candidates = FILTER cost_data BY keyword_list	Find potential matches
6	filtered = APPLY_EXCLUSIONS(candidates, [“underfloor”, “special”])	Remove inappropriate items
7	best_match = SELECT_BEST_FIT(filtered, specifications)	Choose optimal cost item
8	STORE_MATCH(component, best_match)	Record component–cost pairing
9	END FOR	Complete matching process
Cost Calculation Algorithm
1	direct_cost_total = 0	Initialize accumulator
2	FOR EACH matched_component IN component_list	Process each matched item
3	base_mat_cost = matched_component[“Material_Cost”]	Retrieve base material cost
4	base_lab_cost = matched_component[“Labor_Cost”]	Retrieve base labor cost
5	base_eqp_cost = matched_component[“Equipment_Cost”]	Retrieve base equipment cost
6	adj_mat = base_mat_cost × material_multiplier	Apply location adjustment
7	adj_lab = base_lab_cost × labor_multiplier	Apply location adjustment
8	adj_eqp = base_eqp_cost × equipment_multiplier	Apply location adjustment
9	unit_cost = adj_mat + adj_lab + adj_eqp	Calculate adjusted unit cost
10	extended_cost = unit_cost × quantity	Multiply by component quantity
11	direct_cost_total += extended_cost	Accumulated to direct total
12	END FOR	Complete cost calculation

, the first Algorithm handles database integration and location modifier retrieval. The second Algorithm performs intelligent component–cost matching using keyword filtering and exclusion logic. The third Algorithm calculates adjusted unit costs and extended costs for all components. These algorithms are executed in sequence by the LLM orchestrator, with outputs from earlier algorithms serving as inputs to subsequent ones.

Location modifier database processing loads Craftsman Zip Area Modifiers 2025.xlsx containing 40,777 ZIP code records. For project ZIP code 01003, the system retrieves the following: material factor: +5% (indicating materials cost 5% above national average), labor factor: +36% (reflecting Boston area’s significantly higher labor rates), equipment factor: +2% (minimal equipment cost variation), weighted average: +19% (composite adjustment across all cost categories). These modifiers reflect regional market conditions. Boston’s high labor factor (+36%) aligns with the city’s elevated wages, union labor rates, and cost of living. The modest material factor (+5%) suggests competitive material markets with moderate logistics costs. Equipment costs show minimal variation (+2%), indicating relatively standardized national equipment rental markets. The system executed the first Algorithm (Table 3) to retrieve location modifiers for ZIP code 01003, resulting in the following multipliers: Material = 1.05, Labor = 1.36, Equipment = 1.02. These values were then passed to the third Algorithm for application during cost calculation.

4.2.3. Phase 3: Component Matching

Receptacle matching illustrates the intelligent component–cost matching algorithm. The AI system formulates search criteria: primary category keyword: “receptacles”, secondary description filter: “duplex 20 amp” (matching standard specifications), exclusion filters: [“underfloor”, “floor box”, “special”] (eliminating specialty types). A database search returns filtered candidates. The initial category filter produces 247 items containing “receptacles”. Description refinement narrows to 15 items with “duplex” and “20 amp”. Exclusion filtering reduces to five candidate items, eliminating specialty installations. Best-fit selection chooses the following: category: “Receptacles in 1-1/2” deep handy boxes with handy box cover”, description: “Duplex 20 amp brown”, base material: USD 24.00 per EA, base labor: USD 17.10 per EA, Base Equipment: USD 0.00 per EA, Unit: EA (each). This selection represents standard construction practice for commercial receptacle installations with appropriate material and labor cost profiles. Conduit matching follows a similar logic with size-specific criteria. Search parameters include the following: primary category: “EMT conduit”, size specification: “3/4” (matching BIM model conduit diameter), installation context: “floor slab or multiple runs” (appropriate for typical installations). Selected match: category: “EMT conduit in floor slab or multiple runs on a trapeze”, description: “3/4”, base material: USD 76.10 per CLF, base labor: USD 157.00 per CLF, base equipment: USD 0.00 per CLF, unit: CLF (cost per hundred linear feet). Panel matching incorporates capacity specifications. search criteria: primary category: “Indoor 120/240-volt C.B. panelboards”, capacity requirements: “125A” (amperage), “8 poles” (circuit positions). Selected match: category: “Indoor 120/240-volt C.B. panelboards (NQO), main lugs only”, description: “125A 8 poles”, base material: USD 937.00 per EA, base labor: USD 153.00 per EA, base equipment: USD 0.00 per EA, unit: EA. All three component types achieved 100% matching success with appropriate cost items from the database, demonstrating the robustness of the AI-powered matching algorithm.

4.2.4. Phase 4: Cost Calculations

Line-item calculations demonstrate the complete cost computation workflow.

For receptacles:

Material: USD 24.00/EA, Labor: USD 17.10/EA, Equipment: USD 0.00/EA.

Location modifier application:

$$\mathrm{A}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{M}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{l}=\mathrm{\$}24.00\times 1.05=\mathrm{\$}25.20/\mathrm{E}\mathrm{A}$$

$$\mathrm{A}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{L}\mathrm{a}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{r}=\mathrm{\$}17.10\times 1.36=\mathrm{\$}23.26/\mathrm{E}\mathrm{A}$$

$$\mathrm{A}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{E}\mathrm{q}\mathrm{u}\mathrm{i}\mathrm{p}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}=\mathrm{\$}0.00\times 1.02=\mathrm{\$}0.00/\mathrm{E}\mathrm{A}$$

$$\mathrm{U}\mathrm{n}\mathrm{i}\mathrm{t} \mathrm{c}\mathrm{o}\mathrm{s}\mathrm{t}:\mathrm{\$}25.20+\mathrm{\$}23.26+\mathrm{\$}0.00=\mathrm{\$}48.46/\mathrm{E}\mathrm{A}$$

Extended cost (30 receptacles):

$$\mathrm{M}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{l}:\mathrm{\$}25.20\times 30=\mathrm{\$}756.00$$

$$\mathrm{L}\mathrm{a}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{r}:\mathrm{\$}23.26\times 30=\mathrm{\$}697.80$$

$$\mathrm{E}\mathrm{q}\mathrm{u}\mathrm{i}\mathrm{p}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}:\mathrm{\$}0.00\times 30=\mathrm{\$}0.00,\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}:\mathrm{\$}48.46\times 30=\mathrm{\$}1453.80$$

For conduit (24.50 CLF):

Material: USD 76.10/CLF, Labor: USD 157.00/CLF, Equipment: USD 0.00/CLF.

Location modifier application:

$$\mathrm{A}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{M}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{l}=\mathrm{\$}76.10\times 1.05=\mathrm{\$}79.91/\mathrm{C}\mathrm{L}\mathrm{F}$$

$$\mathrm{A}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{L}\mathrm{a}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{r}=\mathrm{\$}157.00\times 1.36=\mathrm{\$}213.52/\mathrm{C}\mathrm{L}\mathrm{F}$$

$$\mathrm{A}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{E}\mathrm{q}\mathrm{u}\mathrm{i}\mathrm{p}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}=\mathrm{\$}0.00\times 1.02=\mathrm{\$}0.00/\mathrm{C}\mathrm{L}\mathrm{F}$$

$$\mathrm{U}\mathrm{n}\mathrm{i}\mathrm{t} \mathrm{c}\mathrm{o}\mathrm{s}\mathrm{t}:\mathrm{\$}79.91+\mathrm{\$}213.52+\mathrm{\$}0.00=\mathrm{\$}293.43/\mathrm{C}\mathrm{L}\mathrm{F}$$

Extended cost:

$$\mathrm{M}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{l}:\mathrm{\$}79.91\times 24.50=\mathrm{\$}1957.80$$

$$\mathrm{L}\mathrm{a}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{r}:\mathrm{\$}213.52\times 24.50=\mathrm{\$}5231.24$$

$$\mathrm{E}\mathrm{q}\mathrm{u}\mathrm{i}\mathrm{p}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}:\mathrm{\$}0.00\times 24.50=\mathrm{\$}0.00$$

$$\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}:\mathrm{\$}293.43\times 24.50=\mathrm{\$}7189.04$$

For panels (two units):

Material: USD 937.00/EA, Labor: USD 153.00/EA, Equipment: USD 0.00/EA.

Location modifier application:

$$\mathrm{A}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d}\mathrm{M} \mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{l}=\mathrm{\$}937.00\times 1.05=\mathrm{\$}983.85/\mathrm{E}\mathrm{A}$$

$$\mathrm{A}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{L}\mathrm{a}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{r}=\mathrm{\$}153.00\times 1.36=\mathrm{\$}208.08/\mathrm{E}\mathrm{A}$$

$$\mathrm{A}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{E}\mathrm{q}\mathrm{u}\mathrm{i}\mathrm{p}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}=\mathrm{\$}0.00\times 1.02=\mathrm{\$}0.00/\mathrm{E}\mathrm{A}$$

$$\mathrm{U}\mathrm{n}\mathrm{i}\mathrm{t} \mathrm{c}\mathrm{o}\mathrm{s}\mathrm{t}:\mathrm{\$}983.85+\mathrm{\$}208.08+\mathrm{\$}0.00=\mathrm{\$}1191.93/\mathrm{E}\mathrm{A}$$

Extended cost (two panels):

$$\mathrm{M}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{l}:\mathrm{\$}983.85\times 2=\mathrm{\$}1967.70$$

$$\mathrm{L}\mathrm{a}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{r}:\mathrm{\$}208.08\times 2=\mathrm{\$}416.16$$

$$\mathrm{E}\mathrm{q}\mathrm{u}\mathrm{i}\mathrm{p}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}:\mathrm{\$}0.00\times 2=\mathrm{\$}0.00$$

$$\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}:\mathrm{\$}1191.93\times 2=\mathrm{\$}2383.86$$

Aggregate direct costs:

$$\mathrm{M}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{l} \mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}:\mathrm{\$}756.00+\mathrm{\$}1957.80+\mathrm{\$}1967.70=\mathrm{\$}4681.50$$

$$\mathrm{L}\mathrm{a}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{r} \mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}:\mathrm{\$}697.80+\mathrm{\$}5231.24+\mathrm{\$}416.16=\mathrm{\$}6345.20$$

$$\mathrm{E}\mathrm{q}\mathrm{u}\mathrm{i}\mathrm{p}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t} \mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}:\mathrm{\$}0.00+\mathrm{\$}0.00+\mathrm{\$}0.00=\mathrm{\$}0.00$$

$$\mathrm{D}\mathrm{i}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t} \mathrm{C}\mathrm{o}\mathrm{s}\mathrm{t} \mathrm{S}\mathrm{u}\mathrm{b}\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}:\mathrm{\$}4681.50+\mathrm{\$}6345.20+\mathrm{\$}0.00=\mathrm{\$}\mathrm{11,026.70}$$

Markup application follows industry-standard sequential calculation:

$$\mathrm{I}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t} \mathrm{c}\mathrm{o}\mathrm{s}\mathrm{t}\mathrm{s}(15\% \mathrm{o}\mathrm{f} \mathrm{d}\mathrm{i}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t}):\mathrm{\$}\mathrm{11,026.70}\times 0.15=\mathrm{\$}1654.01$$

$$\mathrm{S}\mathrm{u}\mathrm{b}\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{w}\mathrm{i}\mathrm{t}\mathrm{h} \mathrm{I}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t}:\mathrm{\$}\mathrm{11,026.70}+\mathrm{\$}1654.01=\mathrm{\$}\mathrm{12,680.71}$$

$$\mathrm{P}\mathrm{r}\mathrm{o}\mathrm{f}\mathrm{i}\mathrm{t}(10\% \mathrm{o}\mathrm{f} \mathrm{s}\mathrm{u}\mathrm{b}\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{w}\mathrm{i}\mathrm{t}\mathrm{h} \mathrm{i}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t}):\mathrm{\$}\mathrm{12,680.71}\times 0.10=\mathrm{\$}1268.07$$

$$\mathrm{S}\mathrm{u}\mathrm{b}\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{w}\mathrm{i}\mathrm{t}\mathrm{h} \mathrm{P}\mathrm{r}\mathrm{o}\mathrm{f}\mathrm{i}\mathrm{t}:\mathrm{\$}\mathrm{12,680.71}+\mathrm{\$}1268.07=\mathrm{\$}\mathrm{13,948.78}$$

$$\mathrm{S}\mathrm{a}\mathrm{l}\mathrm{e}\mathrm{s} \mathrm{t}\mathrm{a}\mathrm{x}(0\% \mathrm{o}\mathrm{f} \mathrm{m}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{l}\mathrm{s}):\mathrm{\$}4681.50\times 0.00=\mathrm{\$}0.00$$

$$\mathrm{G}\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{d} \mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}:\mathrm{\$}\mathrm{13,948.78}+\mathrm{\$}0.00=\mathrm{\$}\mathrm{13,948.78}$$

4.2.5. Phase 5: Visualization Generation

The AI system generates a comprehensive HTML dashboard incorporating modern web technologies and professional design principles, as shown in Figure 6. The structure includes semantic HTML5 markup for accessibility and standards compliance. CSS3 styling includes gradient backgrounds, shadow effects, and responsive grid layouts. Interactive features include hover effects, print styling optimization, and dynamic content organization. The executive summary section is displayed prominently at the top, with the gradient background (purple to violet) commanding attention. The grand total cost is shown in large typography (USD 13,945.81), immediately communicating the primary estimate result. Project location (ZIP 01003) and key parameters (15% indirect, 10% profit) provide a quick context. Cost breakdown visualization presents material, labor, and equipment distributions. Component cards show individual system costs: Conduit System: USD 7188.91 (65.2% of direct costs), Receptacles: USD 1453.68 (13.2% of direct costs), Electrical Panels: USD 2381.76 (21.6% of direct costs). Color-coded boxes use green for material costs, blue for labor costs, and orange for equipment costs, creating visual distinctions between cost categories. Line-item estimate table provides complete transparency with columns for Description, Quantity, Unit, Material, Labor, and Total Cost. Each of the three-line items (receptacles, conduit, and panels) is fully detailed. Subtotals show direct costs before markups.

Markup sections clearly present indirect costs (15%), profit (10%), and tax (0%) calculations with percentage rates and dollar amounts. The cost distribution analysis section shows percentage breakdowns by system with visual bar charts. Key cost drivers are identified: Conduit installation represents the largest component at 65.2%, driven by a high labor content and substantial linear footage. Labor costs dominate at 57.5% of direct costs, reflecting Boston’s +36% labor location factor. Material costs comprise 42.5% of direct costs with +5% regional adjustment. Notes and assumptions section documents important parameters include the ZIP code 01003 location factors applied; the Craftsman 2025 pricing basis; indirect costs at 15%, profit at 10%, and sales tax at 0%; exclusions such as permits, inspections, wires/cables, and breakers; and specific inclusions such as conduits, receptacles with boxes, and panels with specified capacity. The print function enables PDF generation through browser print functionality, with CSS media queries optimizing appearance for printed/PDF output by removing interactive elements, adjusting backgrounds for ink conservation, ensuring page break optimization, and maintaining professional appearance in static format.

5. Results and Discussion

This section presents the empirical outcomes of the automated cost estimation case study, followed by a comprehensive validation against manual benchmarks. The results are analyzed to evaluate the system’s accuracy, efficiency, and reliability.

5.1. Quantity Extraction Accuracy Assessment

The AI-driven workflow processed the electrical system of the commercial building model, executing the five-phase sequence without human intervention. The initial output, generated in under 45 s, provided a complete cost breakdown. The system extracted 187 BIM elements and matched them to appropriate cost items from the Craftsman 2025 database. Using the location modifiers for ZIP code 01003 (Material: +5%, Labor: +36%, Equipment: +2%), the AI calculated a preliminary grand total of USD 13,945.81. This result included all material, labor, and equipment costs, with sequential application of 10% indirect costs and 10% profit margin. Automated extraction identified 30 receptacles; manual verification confirmed 29 receptacles. Discrepancy: one receptacle (3.4% overcount). This variance resulted from the AI counting a junction box or disconnect as a receptacle due to similar BIM category classification.

Automated extraction identified 155 conduit segments totaling 2450 linear feet; manual verification confirmed 156 conduit segments totaling 2526.5 linear feet (25.265 CLF). Discrepancy: one additional segment and 76.5 additional linear feet (3.1% length undercount). The automated system appears to have aggregated a connected conduit run into single segments, missing junctions and direction changes that create distinct elements in Revit. Both automated extraction and manual verification confirmed two electrical distribution panels at 125 amp, 20-pole capacity (100% accuracy match). These findings demonstrate that while the MCP-based extraction achieves high accuracy for discrete equipment items (panels), it requires refinement for complex geometric elements (conduit networks) and may misclassify similar element types (receptacles vs. junctions).

Corrected Cost Calculation Results:

Using manually verified quantities, the corrected cost estimate is as follows:

• Direct total costs: USD 11,205.02, comprising the following:
• Material: USD 4739.72 (42.3% of direct costs);
• Labor: USD 6465.31 (57.7% of direct costs);
• Equipment: USD 0.00 (0.0% of direct costs);
• Indirect costs (15%): USD 1680.78;
• Profit (10% on subtotal with indirect): USD 1120.52;
• Sales tax (0%): USD 0.00;
• Grand total: USD 14,006.28.

5.2. Confidence Intervals and Professional Accuracy Standards

Industry-standard accuracy expectations for construction cost estimates vary by project phase and detail level. For detailed estimates based on complete design, typical accuracy ranges are ±5–10% at 90% confidence or ±8–12% at 95% confidence [33]. The automated system produced a −5.1% variance from the manually verified estimate. This result falls within acceptable ranges for preliminary estimates (±10–15%) but exceeds tolerance for detailed estimates (±5–8%) [33].

Aggregate Estimate Confidence (Using Corrected Quantities):

• Component-level variance expectations with corrected quantities:

At 95% confidence interval:

• Overall variance: ±8% USD 14,006.28× 0.08 = USD 1120.50);
• Estimated range: USD 12,885.78 to USD 15,126.78.
• This range encompasses typical detailed estimate uncertainty.

At 90% confidence interval:

• Overall variance: ±6% (USD 14,006.28× 0.06 = USD 840.38);
• Estimated range: USD 13,165.90 to USD 14,846.66.
• This is a more conservative range for bid-level accuracy.

The corrected estimate with manual quantity verification achieves professional accuracy standards. However, the automated extraction alone (without manual verification) produced a −5.1% variance, indicating that the current MCP-based extraction requires quality control validation for production use.

5.3. Database Search and Matching Accuracy Assessment

The cost database search and component matching algorithms demonstrated 100% accuracy in retrieving appropriate cost items from the Craftsman National Building Cost Manual 2025, as shown in Figure 7. Once component types were identified (receptacles, conduit, and panels), the AI-powered matching system successfully located corresponding cost items with correct specifications. All matched items contained complete cost information with no missing values or incorrect category selections. The location modifier database queries similarly achieved 100% accuracy, correctly retrieving ZIP code 01003 modifiers (+5% material, +36% labor, and +2% equipment) and applying them appropriately to all cost calculations.

5.4. Discussion

The results of the case study show the potential transformation of merging LLMs with BIM using standardized protocols. The main goal is to create a pipeline where a conversational AI agent can conduct a cost estimating workflow using the MCP autonomously. This result differs from earlier automation initiatives, which focused on individual estimating steps or required heavy exclusive integration. The 98.6% decrease in processing time, from hours to seconds, confirms that AI orchestration can automate quantification and costing procedures. This efficiency improvement requires quality assurance, especially for complicated geometric parts, as can be seen by the extraction accuracy discrepancy.

The work shows how the MCP efficiently addresses the “M × N” integration problem in building software ecosystems, making a remarkable methodological contribution. In contrast to bespoke API interfaces or proprietary plugins, the MCP provides a standardized, bidirectional communication layer that lets the AI find and use features from other tools. This design allows Claude AI, Autodesk Revit, and Excel-based cost databases to communicate without low-level scripting. By integrating new data sources or tools via MCP servers instead of rewriting connection frameworks, this strategy saves long-term maintenance and boosts adaptability. This discovery overcomes a major literary hurdle to AI integration in construction scalability and sustainability.

Empirical validation shows that LLM-driven building automation has a complex performance profile. The system performed 100% cost database matching and financial computations, demonstrating good semantic comprehension, rule-based reasoning, and numerical computing skills. The AI detected and linked the 3/4” EMT conduit from the BIM model to the Craftsman database record, adding correct position modifiers and markup sequences. This outcome implies that LLMs can manage construction estimating domain-specific knowledge and procedural logic. Geometric extraction was less accurate (96.6–100%), making it difficult to measure networked systems such as conduit lines. AI systems are more dependable for interpretative and computational tasks than for exact spatial measurement from complicated BIM geometries, highlighting areas where human validation is still needed.

The −5.1% cost variance before quantity verification affects implementation. This variation surpasses bid-level accuracy but is acceptable for early-design estimates. This outcome proves the system’s best use: quick preliminary estimates that require professional verification for final deliverables. The hybrid workflow model indicates that AI provides baseline estimates at unprecedented speed, freeing up human estimators to focus on validation, value engineering, and sophisticated assembly. Using AI’s speed, consistency, and calculation and human experts’ judgment, expertise, and complicated problem-solving, this collaborative approach may improve estimate quality and efficiency.

This research directly addresses the critical gaps identified in the literature. In contrast to previous approaches that relied on custom, brittle integrations, the novel application of the MCP provides a standardized, scalable solution to the longstanding interoperability problem. Furthermore, while existing tools and research primarily automate isolated tasks like quantity take-off, this study demonstrates the first empirically validated, end-to-end automated workflow orchestrated by an LLM. The system’s performance, reducing estimation time from hours to seconds while maintaining professional accuracy, provides the quantitative evidence needed to advance AI adoption in construction, a gap highlighted by prior reviews. Finally, by detailing a replicable framework and validation protocol, this work moves beyond a proof of concept to offer a practical blueprint for the industry, addressing the replicability gap common in earlier research. The novel contributions of this study are clarified through a direct comparison with established challenges in the field, as summarized in

Table 4. Comparison of the MCP-based framework with the current literature and practice.

Aspect	Current Literature/Practice	Advancement in This Study
Integration Method	Relies on custom, one-off solutions (APIs, plugins) that are hard to scale.	Introduces standardized protocol MCP as a universal adapter, solving the “M × N” integration problem.
Automation Scope	Mostly semi-automated, focusing only on quantity take-off.	Achieves full end-to-end automation from BIM interrogation to final report in <45 s.
AI Application	Focuses on predictive models or theoretical LLM frameworks.	Empirically validates an LLM for domain-specific reasoning, with 100% accuracy in cost matching.
Solution Replicability	Often presents proof-of-concept or proprietary tools without a clear adoption path.	Provides a detailed, replicable blueprint (architecture, algorithms, validation protocol).
Performance Validation	Lacks comprehensive quantitative benchmarks against manual standards.	Delivers rigorous metrics: 98.6% time reduction and accuracy within professional confidence intervals (±8%).

.

Collectively, these contrasts demonstrate that employing a standardized integration protocol (MCP) is the pivotal innovation that enables reliable, LLM-driven, end-to-end automation for construction cost estimation.

This research’s shortcomings provide future research opportunities. While providing depth of research, the single-discipline, single-project case study restricts generalizability across construction domains and project scales. Without validation, electrical system performance may not transfer to mechanical, plumbing, or structural estimates. The research also used static cost databases; real-time price feeds might improve practical implementations. The geometric extraction issues imply that MCP server development for BIM systems should enhance network element quantification and element type disambiguation. A multi-disciplinary case study, scalability testing on bigger projects, and more advanced geometric analysis techniques in the MCP framework should examine these limitations.

5.5. Comparative Analysis with the Existing Literature

To contextualize the findings within the broader body of BIM-based and AI-driven cost estimation research,

Table 5. Comparative performance analysis with prior BIM-based cost estimation studies.

Study	Approach	Automation Level	Time Efficiency	Accuracy Metric
This Study	MCP + LLM	Full end-to-end	98.6% reduction	−5.1% variance
Alazawy et al. (2024) [34]	SVM-BIM	QTO only	Not reported	0.41–1.48% variance
Wang et al. (2022) [28]	DNN	Cost prediction	N/A	MAPE varies by model
Banihashemi et al. (2022) [21]	ML + 5D BIM	Classification + QTO	Significant	Project-dependent
Elmousalami (2020) [27]	Various ML	Prediction only	N/A	7–15% MAPE
Traditional BIM QTO	Manual matching	QTO only	Baseline	Professional standard

presents a comparative analysis of this study’s results against previously published approaches.

The comparative analysis reveals several significant findings across four key dimensions. First, regarding estimation accuracy, the −5.1% cost variance achieved by the MCP-LLM framework falls within the ±10–15% tolerance acceptable for preliminary estimates, though it exceeds the tighter ±5–8% tolerance required for detailed bid-level estimates [33]. This performance is comparable to the 7–15% MAPE reported by Elmousalami [27] for conventional ML models, while offering the additional capability of complete workflow automation rather than prediction alone.

Second, concerning the level of automation, prior studies have predominantly focused on automating isolated tasks within the estimation workflow. Alazawy et al. [34] demonstrated high accuracy (0.41–1.48% variance) for quantity take-off using SVM-BIM integration, but their approach addressed only the quantity extraction phase without subsequent cost matching or report generation. Similarly, Banihashemi et al. [21] developed a machine learning-integrated 5D BIM prototype that classifies building cost elements but still requires human intervention for final estimate compilation. The present study achieves what the literature has identified as a critical gap: full end-to-end automation from model interrogation through professional report generation [17,31].

Third, the efficiency gains documented in this study (98.6% time reduction) exceed those reported in comparative BIM studies. Attia [35] reported that AI-BIM integration achieved 30% efficiency improvement in project workflows, while case studies have documented a 60% faster design time with AI-BIM integration. The substantially higher efficiency in this study reflects the comprehensive automation of the entire estimation pipeline rather than individual workflow segments.

Fourth, regarding semantic reasoning capability, conventional ML approaches require extensive labeled datasets for training and cannot adapt to novel component types without retraining [27]. The LLM-based approach demonstrated 100% accuracy in cost database matching by leveraging semantic reasoning to interpret diverse BIM naming conventions, a capability that addresses the “semantic gap” challenge identified by [15,36] as a fundamental limitation of ontology-based approaches.

These comparisons demonstrate that while the MCP-LLM framework does not achieve the highest accuracy for individual tasks (SVM-BIM achieves tighter QTO variance), it uniquely delivers comprehensive workflow automation with an accuracy sufficient for preliminary estimation and decision support applications.

6. Conclusions

This research presented, implemented, and validated a novel framework for the fully automated estimation of construction costs by integrating Building Information Modeling with large language models through the Model Context Protocol. The study addressed three research questions concerning interoperability (RQ1), accuracy (RQ2), and efficiency (RQ3) through a rigorous design science research methodology.

Addressing RQ1 (Interoperability): The MCP-based architecture successfully bridged the integration gap between Autodesk Revit, the Claude AI engine, and the Craftsman cost database. The protocol enabled dynamic capability discovery and bidirectional data exchange without custom API development, demonstrating that standardized protocols can transform the traditional “M × N integration problem” into a manageable “M + N” configuration. This finding directly addresses the fragmentation challenge identified in the prior literature (Kocakaya et al., 2025 [31]; Ayyagari, 2025 [9]).

Addressing RQ2 (Accuracy): The system achieved 100% accuracy in semantic cost database matching and financial calculations, demonstrating that LLMs possess sufficient domain reasoning capability for construction cost estimation tasks. Quantity extraction accuracy ranged from 96.6% to 100% depending on component complexity, with networked elements (conduits) showing lower accuracy than discrete items (panels). The overall cost variance of −5.1% falls within acceptable tolerances for preliminary estimates (±10–15%) as per industry standards, though professional validation remains necessary for bid-level accuracy (±5–8%).

Addressing RQ3 (Efficiency): The framework achieved a 98.6% reduction in processing time, compressing a workflow conventionally requiring 2.5–3.5 h of manual effort into under 45 s of automated processing. This efficiency gain substantially exceeds the 30–60% improvements reported in comparative AI-BIM integration studies (Attia, 2025 [35]), attributable to the comprehensive automation of all workflow phases rather than individual task optimization.

The key empirical finding is the profound efficiency gain achieved: a task conventionally requiring 2.5–3.5 h of manual effort was reduced to under 45 s of automated processing, representing a time reduction exceeding 98%. This acceleration unlocks the potential for real-time cost feedback during iterative design phases. The system demonstrated flawless performance in semantic tasks, achieving 100% accuracy in cost database matching and financial calculations. However, the study also provided a crucial, realistic assessment of current limitations, identifying variable accuracy (96.6–100%) in the automated geometric extraction of quantities from BIM, particularly for networked elements such as conduit runs. This assessment resulted in an initial cost variance of −5.1%, which was corrected through targeted manual verification.

These findings lead to the central practical conclusion that the most effective application of this technology is not as a replacement for human estimators but as the core of a hybrid human–AI collaborative workflow. In this model, the AI generates rapid, preliminary estimates, freeing skilled professionals to focus their expertise on validation, value engineering, and complex judgment calls. This synergy leverages the unique strengths of both: the AI’s speed, consistency, and computational power, and the human’s contextual understanding, experience, and ability to manage exceptions.

The primary contributions of this work are threefold. Methodologically, it establishes the MCP as a replicable and scalable template for connecting AI to the complex ecosystem of construction software, moving beyond one-off custom integrations. Empirically, it provides quantified evidence of the dramatic efficiency gains possible through AI orchestration while honestly delineating the current boundaries of automation accuracy. Practically, it delivers a functional blueprint and validation protocol that industry practitioners can adapt, lowering the barrier to adoption for AI-driven automation in construction firms.

For future research, several pathways emerge. The immediate priority is the refinement of MCP server capabilities for BIM platforms to improve the geometric reliability of quantity take-off for complex systems. Subsequent work should expand validation across multiple construction disciplines and project scales, develop methods for integrating real-time market data, and explore the AI’s role in automating change detection and quantity tracking throughout the project lifecycle.

In summary, this research confirms that the convergence of LLMs, standardized protocols such as the MCP, and BIM data represents a transformative shift for construction cost management. While not yet autonomous, the developed framework provides a powerful, accurate, and immensely efficient assistant that can reshape estimating workflows. It marks a significant step toward an intelligent, integrated, and data-driven future for construction, where technology amplifies human expertise to achieve unprecedented levels of productivity and precision.