Comprehensive searches were conducted across four major databases: Scopus, ScienceDirect, Google Scholar, and the University of East London Library Portal. Boolean operators combined keywords related to AI, construction scheduling, risk management, and socio-technical integration. Example queries included:

(“Artificial Intelligence” OR “Machine Learning” OR “Genetic Algorithm” OR “Reinforcement Learning”) AND (“Construction Scheduling” OR “Project Planning”)

To ensure the review captured high-quality, scientifically rigorous evidence relevant to the research question, strict eligibility boundaries were established.

Inclusion Criteria:

Document Type: Peer-reviewed journal articles and high-impact conference proceedings were prioritized to ensure the academic rigor of the theoretical data.

Exclusion Criteria:

Grey Literature: Non-peer-reviewed sources such as editorials, opinion pieces, and general web articles were excluded to maintain the “scholarly quality” of the evidence base, with the exception of official government reports (e.g., NAO, IPA) used solely for policy context (Page and al., 2021).

The selection process followed the PRISMA 2020 (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) protocol to ensure transparency and reproducibility (Page et al., 2021). The process moved through four distinct stages: Identification, Screening, Eligibility, and a formal Quality Appraisal.

Stage 1: Identification and deduplication

Initial database searches yielded 1,245 records. These were imported into reference management software where 210 duplicates were automatically removed.

Stage 2: Title and abstract screening

The remaining 1,035 records were screened based on titles and abstracts. Studies clearly outside the scope (e.g., AI in manufacturing or healthcare) were discarded, resulting in 890 exclusions.

Stage 3: Eligibility and quality appraisal (QA)

The remaining 145 full-text articles were assessed for eligibility. To address the need for methodological rigour, a specific Quality Appraisal (QA) step was applied at this stage. Articles were evaluated against three core criteria derived from standard critical appraisal checklists (e.g., CASP):

Methodological Clarity: Does the study clearly define the AI inputs, algorithms, and validation metrics used?

Studies failing to meet these quality thresholds, such as those presenting purely speculative concepts without logical architecture, were excluded.

Stage 4: Final inclusion

This rigorous process resulted in the final selection of 60 high-quality studies for qualitative synthesis. These papers formed the evidence base for diagnosing the “integration gap” and designing the conceptual framework.

Metaheuristic algorithms are widely used to tackle complex scheduling problems by searching vast solution spaces for near-optimal plans. Two leading techniques are Genetic Algorithms (GA) and Particle Swarm Optimisation (PSO) (Abioye et al., 2021). GA, inspired by natural evolution, excels at multi-constraint problems such as the Resource-Constrained Project Scheduling Problem (RCPSP) and Time-Cost Trade-off (TCT), finding global optima while avoiding local traps (Slowík and Kwaśńicka, 2020; Xie et al., 2021). PSO, modelled on bird flocking behaviour, is effective for multi-objective optimisation, balancing time, cost, and quality (Elbeltagi et al., 2016). Despite their strengths, these methods are static - they optimise fixed inputs and cannot adapt to real-time changes, limiting their use to upfront planning rather than dynamic control (Abioye et al., 2021; Shodunke, 2025). In simple terms, static optimisation is based on finding the best initial plan. The evolution of AI paradigm in construction scheduling is presented in Figure 4.

Machine Learning (ML) addresses a key weakness of traditional scheduling: reliance on static, often inaccurate duration estimates. Instead of generating schedules, ML improves input quality through predictive analytics (Abioye et al., 2021). Supervised models, such as deep neural networks, learn complex relationships between project features - task dependencies, resources, weather, and design complexity-and actual durations (Shodunke, 2025). This enables dynamic, probabilistic forecasts, replacing guesswork with evidence-based predictions. Accuracy gains are significant: traditional methods achieve ∼41%, while ML models reach up to 91.4%. Integration with Building Information Modelling (BIM) is critical, as BIM provides structured data for training models to predict durations, sequences, and resource needs directly from design models (Abioye et al., 2021).

Reinforcement Learning (RL) introduces a transformative shift in construction scheduling by enabling dynamic, adaptive control. Unlike Genetic Algorithms (GA) and Particle Swarm Optimisation (PSO), which optimize static plans, RL maintains optimality in real time, making it ideal for uncertain, fast-changing site conditions (Abioye et al., 2021). In this framework, an intelligent agent interacts with a simulated environment, often a Digital Twin, learning through trial and error and refining its decisions based on rewards and penalties (Pregnolato, 2023). This capability allows schedules to self-adjust in response to disruptions such as material delays or adverse weather, marking a move from rigid planning to agile, uncertainty-aware project management (Shodunke, 2025).

The cornerstone of AI-driven risk management is Machine Learning for predictive analytics. Supervised models, particularly Artificial Neural Networks (ANNs), are trained on historical datasets linking project characteristics to risk outcomes such as cost overruns, schedule delays or safety incidents (Abioye et al., 2021). This quantitative, data-driven approach replaces the subjective assessments typical of traditional risk registers (Khodabakhshian, 2023). Once trained, these models predict the probability and severity of risks in new projects, providing an objective basis for prioritising mitigation efforts.

Generative AI further enhances these capabilities. Unlike conventional models that analyse existing data, Generative AI devises adaptive mitigation strategies and generates novel predictive insights (Darko et al., 2020). Operating on continuous learning, these models improve accuracy and strategic effectiveness over time (Patel, 2023). Recent studies comparing advanced Generative AI models such as GPT-4 with human experts show AI producing more comprehensive risk plans, identifying a broader range of risks. While these plans may lack project-specific detail, they highlight the potential for synergy: AI provides a robust baseline, refined by human expertise (Mohamed, 2025).

In the construction domain, critical risk intelligence is frequently embedded within unstructured text sources, including progress reports, contracts, Requests for Information (RFIs), and accident narratives, rather than structured databases (Jagannathan et al., 2022). Consequently, Natural Language Processing (NLP) is specifically deployed to automate the extraction of this linguistic data. Key applications include the scanning of legal and contractual documents to flag ambiguous or high-risk clauses that historically precipitate disputes (Hassan et al., 2021). Furthermore, NLP algorithms analyse safety reports and accident narratives to detect recurring patterns and root causes, thereby facilitating proactive prevention strategies rather than retrospective analysis.

Computer Vision (CV) brings real-time visual intelligence to construction safety management, enabling the continuous analysis of video streams from site cameras or drones to detect hazards instantly (Abioye et al., 2021). Leveraging deep learning models such as Convolutional Neural Networks (CNNs) and object detection algorithms like You Only Look Once (YOLO), these systems monitor specific site risks, including compliance with Personal Protective Equipment (PPE) policies, fall hazards, and unsafe proximity to heavy machinery. While real-world deployments have significantly reduced near-miss incidents by providing non-intrusive oversight beyond the capacity of human managers, the technology is not infallible (Alateeq et al., 2023). It is critical to note that CV systems remain prone to false positives caused by environmental factors such as variable site lighting or visual occlusion. Consequently, to ensure reliability and prevent erroneous alerts, these automated systems necessitate the integration of human verification loops.

Building Information Modelling (BIM) serves as the primary repository for managing information across a project’s lifecycle. As a digital representation of a facility’s physical and functional characteristics, BIM delivers the structured, object-oriented data required for training and deploying AI models. Geometric and non-geometric data from BIM underpin AI applications such as automated cost estimation, clash detection, and quality monitoring.

While the UK government mandates BIM through the UK BIM Framework to drive standardisation, interoperability remains a persistent technical hurdle. The fragmentation of proprietary software often results in data loss during export to open standards like IFC, creating barriers to the seamless data flow required for AI integration (Ghaffarianhoseini et al., 2017; Sacks et al., 2020; Abioye et al., 2021).

Digital Twins extend BIM by creating dynamic, virtual models of physical assets that update continuously with real-time sensor data (Abioye et al., 2021). This live cyber-physical link enables simulation, prediction and optimisation of performance (Sjödin et al., 2020). Digital Twins are critical for advanced AI applications: they provide the real-time data streams needed for predictive risk analytics and serve as the simulated environment for training Reinforcement Learning agents in adaptive scheduling (Callcut et al., 2021). By allowing AI to interact with a data-rich virtual replica, Digital Twins make real-time autonomous schedule control feasible.

Despite their theoretical benefits and strategic importance in the UK’s National Digital Twin Programme, significant barriers to widespread adoption remain (Trade, 2023). Cost is a primary constraint, particularly regarding the deployment of IoT sensor networks and cloud infrastructure, which may be prohibitive for Small and Medium Enterprises (SMEs). Furthermore, data governance presents complex challenges; specifically, unresolved questions regarding data ownership, security protocols, and liability when sharing live asset models across supply chains continue to hinder implementation maturity (Boje et al., 2020).

The Technology Acceptance Model (TAM) explains individual adoption of new technology through two key beliefs: perceived usefulness and perceived ease of use (Davis, 1989; Marikyan and Papagiannidis, 2024). In UK construction, many experienced professionals doubt AI’s ability to outperform their intuition and view AI tools as complex, opaque ‘black boxes’. This scepticism, combined with perceived difficulty, creates strong individual-level resistance to adoption (Jallow et al., 2022).

Socio-Technical Systems Theory asserts that technical systems, such as AI frameworks, and social systems - comprising people, skills, culture and workflows - are fundamentally interdependent. Optimising technology without considering its organisational context is likely to result in failure (Oke et al., 2023). This insight is particularly relevant to the UK construction sector, where adversarial practices, fragmented structures and persistent digital skills gaps pose barriers as significant as any technical challenge (Kelvin Ibrahim and Aliu, 2025). The adoption of Construction 4.0 technologies has far-reaching implications for workforce dynamics, employment models and labour relations, which must be managed proactively. Even when individual managers embrace AI tools—overcoming barriers identified by the Technology Acceptance Model—their organisations may lack the integrated processes or collaborative structures needed to support real-time data use, resulting in abandonment of the technology. A successful integrated framework must therefore be designed with the social system in mind, not merely as a technical solution (Abioye et al., 2021).

Institutional Theory explains industry-level inertia by showing that organisations adopt new practices not only for efficiency but to gain legitimacy through conformity to norms (Kauppi, 2022). This conformity is shaped by three pressures: coercive, mimetic and normative. In UK construction, coercive pressure from government BIM mandates mainly affects large public projects, leaving SMEs largely untouched. Mimetic pressure is the tendency for organisations to copy successful competitors, especially in times of uncertainty. Mimetic pressure is weak in a fragmented industry where innovation is hard to observe. Normative pressure from professional bodies like RICS and ICE diffuses slowly, especially among SMEs (ONS, 2024).The result is a weak institutional environment with no strong imperative for widespread digital adoption.

This core layer is the essential foundation for any AI solution, built to eliminate the severe data integration and interoperability barriers that hinder adoption (Sjödin et al., 2020; Abioye et al., 2021) Without a unified, reliable data backbone, downstream AI engines would produce flawed outputs. In the UK context, it ensures compliance with national standards, including the UK BIM Framework and Common Data Environments (CDEs), aligning with government mandates for data standardisation on public projects (Cabinet Office, 2022). By linking BIM models, ERP systems and IoT sensors, it creates a single source of truth on which all subsequent AI processes depend.

The literature review suggests that, this engine serves as a multi-modal “sensorium,” capturing risk signals across numerical, textual and visual data streams. Its modular design integrates specialised AI models, each acting as a distinct “sense,” to deliver a comprehensive risk profile. It comprises the following sub-components:

Predictive Analytics: ML models such as GBDT achieve up to 87.3% accuracy in risk prediction, outperforming traditional methods by 23% (Khodabakhshian, 2023).

However, it is important to note that these figures represent potential capabilities reported in controlled studies; real-world performance will vary significantly based on data quality and site complexity.

The Risk-to-Constraint Translation Engine acts as the framework’s semantic bridge, designed to close the critical “integration gap” between sensing and scheduling. Acting as an intelligent interpreter, it links the diverse outputs of the Risk Analysis Engine with the structured inputs required by the Dynamic Scheduling Engine. In essence, it converts abstract risk signals, whether probabilities, qualitative indicators, or binary states, into actionable project constraints on time, cost, and resources. It is important to note that this component is currently defined as a high-level logic model that maps risk types to specific constraint classes. The development of the specific computational algorithms and executable pseudo-code required to operationalise this logic is identified as the immediate next step in future research.

Reinforcement Learning (RL) marks a decisive shift from static planning to dynamic control. Unlike traditional algorithms that optimise fixed plans, RL adapts to real-time uncertainty by learning optimal decision sequences in changing conditions (Abioye et al., 2021). Its safe, practical deployment is achieved through integration with a Digital Twin, providing a high-fidelity “sandbox” for thousands of simulations without real-world risk (Pregnolato, 2023).

This user-facing layer tackles the non-technical barriers to adoption in UK construction. It directly addresses the key drivers of acceptance set out in the Technology Acceptance Model (TAM): Perceived Usefulness and Ease of Use (Davis, 1989; Marikyan and Papagiannidis, 2024). The interface boosts usefulness through clear, data-driven insights and enhances ease of use with an intuitive dashboard, making AI accessible and effortless. Crucially, its human-in-the-loop design counters “black box” scepticism (Jallow et al., 2022), ensuring managers remain the ultimate decision-makers and reinforcing trust by augmenting, not replacing, their expertise (Oke et al., 2023).

The use of Computer Vision (CV) to detect unsafe behaviours (e.g., PPE non-compliance) creates a tension between safety enforcement and worker surveillance. Under the UK General Data Protection Regulation (UK GDPR), processing biometric data requires a lawful basis and proportionality. If the CV system is perceived as a tool for invasive performance monitoring rather than protective oversight, it will face resistance consistent with the “job security” fears identified in the anatomy of resistance (Kelvin Ibrahim and Aliu, 2025). Therefore, the framework’s data governance must enforce privacy-by-design-anonymising video feeds so that they detect hazards (e.g., “missing helmet”) rather than identifying individuals, ensuring the technology enhances welfare without violating privacy rights (Alateeq et al., 2023).

The Construction (Design and Management) Regulations CDM 2015 (Great Britain, 2015) place strict legal duties on Principal Contractors to manage health and safety. A critical legal risk arises if the AI system fails to identify a hazard or suggests an unsafe schedule sequence. It must be explicitly clear that the AI functions as a decision-support tool, not a decision-maker. The legal liability under CDM 2015 remains with the human duty holder. The framework’s “Decision-Support Interface” is therefore not just a user experience feature; it is a regulatory necessity to ensure that a human manager reviews and accepts responsibility for the AI’s recommendations, maintaining the chain of accountability required by UK law (HSE, 2025).