In controlled evaluations of construction planning, researchers have demonstrated that concentrating cooperative feedback on the most divided human participants while providing static artificial intelligence references reduces decision-making communication by more than half. The research, published in the journal Applied Intelligence by Hezheng Mao, develops a competence-based multi-mode framework to incorporate artificial intelligence assessments into group decisions regarding complex building plans. Modern building projects require coordinated execution among participants whose specialized responsibilities, professional backgrounds, and personal priorities frequently differ. Autonomous agents can generate complete evaluations of construction options, creating a pressing need to integrate machine perspectives into collective planning without diluting human managerial accountability. In simulated trials evaluating five comprehensive execution packages across two batches of 180 cases, the framework cut required communication events by 50.39 percent and decreased maximum local disagreement by 19.78 percent compared to random-pair cooperation. This approach demonstrates a traceable path for utilizing artificial intelligence advice within human-led engineering consensus.
The operational shift toward intelligent construction connects physical site activities with advanced digital technologies, including building information modeling, digital twins, pervasive sensing networks, and predictive data analytics. Digital twin workflows continuously link ongoing physical observations with information used to control production, generating rich empirical streams for management teams. However, deciding on a single execution plan requires aligning professionals who bear distinct operational burdens. Project managers, site safety officers, structural engineers, and field superintendents regularly evaluate identical options from differing viewpoints. For high-rise building projects, selecting an execution plan demands joint evaluation of work sequences, access logistics, protective installations, and resource allocation. Among feasible plans, planners often weigh safety protections against budget constraints and operational disruptions. Meanwhile, generative artificial intelligence and large language models have progressed from passive sensor alerting to producing multimodal safety inspections, structured risk assessments, and procedural disassembly proposals, demonstrating substantial evaluative capacity for complex building decisions.
This growing technical capability creates a representational gap in standard group decision models, which typically absorb machine findings as background context rather than tracing their explicit deliberative influence. When algorithms are treated merely as unrecorded data feeds, planners cannot systematically audit how specific recommendations alter individual viewpoints. Furthermore, construction management requires clear boundaries between contributing an opinion and accepting responsibility for implementation. While artificial intelligence agents can assess complete execution alternatives, they do not assume equivalent organizational commitments or obligations. The framework resolves this challenge by separating opinion contribution from implementation responsibility. Human planners and computational agents evaluate identical candidate plans, but artificial intelligence evaluations remain fixed throughout each coordination episode. Machines neither negotiate nor alter their submitted matrices. Humans retain complete ownership of the consensus process, determining how outside assessments are incorporated, defining the collective preference center, and executing the final plan selection.
To establish an equitable mathematical foundation for comparing heterogeneous viewpoints, the framework models all assessments using an improved cardinal fuzzy preference relation. In reciprocal preference matrices, an entry exceeding 0.5 indicates a preference for one alternative over another, while 0.5 represents exact indifference. Traditional distance metrics often confound the numerical intensity of a preference with its direct comparative ordering. For instance, values of 0.51 and 0.49 are numerically proximate but support opposite decisions, whereas values of 0.55 and 0.85 endorse the same plan at divergent intensities. The framework addresses this nuance by formulating a composite similarity metric that combines intensity similarity and direction similarity through a balancing parameter. Intensity similarity measures the mean absolute deviation between paired numerical scores across all plan combinations, whereas direction similarity evaluates whether participants agree on the ordinal superiority of each alternative. This dual-component structure ensures that diagnostic similarity captures both detailed quantitative agreement and practical directional consensus.
Because pairwise similarity evaluates alignment between separate participants, the system relies on an independent structural quality index to appraise the internal validity of individual submissions. A submitted matrix must exhibit internal logical consistency and clear alternative differentiation to serve as an effective reference during negotiation. Internal consistency is calculated by extracting unnormalized alternative priorities from pairwise comparison ratios, constructing a multiplicatively consistent benchmark matrix, and measuring the submission’s deviation from this reconstructed ideal. In parallel, normalized information discrimination is computed using Shannon entropy to confirm that priorities are distinctly separated rather than flatly uniform. These two properties are merged using a parameter-two Hamacher t-norm, yielding a composite structural quality score. This mathematical formulation strictly penalizes assessments lacking either logical transitivity or decisive discrimination, ensuring that highly coherent but entirely indifferent opinions receive low structural scores. Maintained statically throughout deliberation, structural quality provides a safeguard against referencing chaotic or non-discriminating preference profiles during subsequent consensus feedback.
The incorporation of advice during deliberation is further moderated by an operational trust structure that couples subjective acceptance with objective task competence. Human participants express raw directed acceptance toward peer planners and artificial intelligence agents, capturing individual willingness to rely on external assessments. To prevent reliance from reflecting unverified assumptions, raw scores are calibrated against external composite competence ratings. For human experts, competence synthesizes educational qualifications, relevant planning experience, professional performance records, and optional physiological readiness indicators. For artificial intelligence agents, competence reflects demonstrated accuracy on comparable tasks, transparency of reasoning, and analytical stability across fluctuating site conditions. The calibration formula scales raw acceptance by an external competence multiplier, anchoring interpersonal trust in verifiable ability. Simultaneously, source support combines received group acceptance with initial structural quality, establishing fixed weights that govern how readily each participant’s opinion is absorbed. These parameters ensure that opinion uptake reflects both verified expertise and the structural coherence of the submitted advice.
Deliberative disagreements are mapped across a partitioned multi-mode matrix that distinguishes human-human, human-machine, and machine-machine comparisons. Machine-machine similarities remain constant throughout coordination, providing an objective diagnostic indicator of whether different algorithms or sensing inputs yield conflicting plan appraisals. Human-machine similarities expose how closely human professionals agree with computational assessments, highlighting points of conceptual divergence across technical domains. To evaluate collective alignment, the framework defines individual agreement as a planner’s mean similarity to other human colleagues, while group consensus is measured by averaging pairwise similarities across the human cohort. Deliberation targets a predefined human consensus threshold of 0.80. Alongside group consensus, the framework monitors maximum local disagreement, defined by the least compatible human pair. This metric ensures that severe bilateral conflicts are not obscured by a satisfactory overall average, preventing scenarios where two key project leaders remain deeply divided over work sequences or access rights while overall statistical consensus appears fulfilled.
When the human consensus level falls short of the target threshold, the framework executes a targeted cooperative feedback mechanism structured into four distinct stages. First, the algorithm identifies the single most divergent human pair, concentrating feedback effort on the sharpest conflict rather than issuing broad revision requests to the entire committee. Second, the system computes the weighted human preference center and selects an external reference source exhibiting positive structural quality that minimizes root-mean-square distance to that center. This reference may be an expert human or an artificial intelligence agent. Third, both selected individuals formulate intermediate mutual compromises, adjusting toward each other based on bilateral acceptance and source support. Fourth, both individuals simultaneously update their evaluations toward the common reference. The common uptake rate is constrained by the minimum of their respective allowances, guaranteeing that neither party is forced to incorporate external advice beyond their willingness. Unselected humans maintain their existing preferences, and machine assessments remain strictly unaltered throughout.
Rigorous mathematical analysis establishes that this two-tier cooperative adjustment guarantees continuous preference difference contraction within each executed feedback step. Because mutual concession fractions and reference absorption parameters represent convex combinations bounded strictly between zero and one, all intermediate and final updates strictly preserve matrix reciprocity, self-indifference, and interior preference bounds. When subtracting the updated matrices of the two participants, identical common reference contributions cancel out algebraically. The remaining matrix difference equals the prior difference multiplied by a scalar contraction factor strictly less than one, ensuring continuous convergence between the selected individuals. The empirical properties of the framework were tested using controlled numerical simulations of five fall-prevention and work-coordination packages across four operational dimensions: exposure-control continuity, resource feasibility, work-interface coordination, and monitoring integration. The experimental design covered two batches of 180 simulation cases across panel cohorts of 6, 12, and 24 humans paired with three static artificial intelligence agents under diverse-preference and minority-trade-off scenarios.
Simulation results confirmed that targeted cooperative feedback achieved substantial efficiency gains across all tested scenarios. In the confirmation batch, targeted cooperation reached the 0.80 consensus threshold with an average of 10.589 feedback events, reducing communication by 50.39 percent compared to 21.344 events under random-pair selection, while reducing cumulative preference adjustment by 18.72 percent and maximum local disagreement by 19.78 percent. Static machine references directly altered human decision trajectories: in 21 of the 180 confirmation cases, an artificial intelligence agent served as the common reference, modifying final aggregated scores in all 21 instances, altering six complete rankings, and changing the top-ranked plan twice. Under common feedback allowances of 140 events, the targeted method drove residual maximum disagreement to 0.0000355, compared to 0.01694 for random pairing. Future research aims to extend this framework to distributed construction sites through hierarchical clustering and natural-language preference processing, offering an accountable mechanism for integrating artificial intelligence into collaborative engineering decisions.
Subject of Research: Human–machine cooperative decision-making and consensus frameworks in intelligent construction planning
Article Title: A competence-based multi-mode framework for human–machine decision-making in intelligent construction
Article References: Mao, H. (2026). A competence-based multi-mode framework for human–machine decision-making in intelligent construction. Applied Intelligence, 56(16), Article 555. https://doi.org/10.1007/s10489-026-07575-6
Image Credits: AI Generated
DOI: 10.1007/s10489-026-07575-6
Keywords: intelligent construction, human-machine decision-making, cooperative feedback, preference diagnosis, group consensus, multi-attribute decision-making, cardinal fuzzy preference relations, structural quality index, competence-based trust, directed acceptance, fall prevention packages, construction execution planning, local disagreement reduction, static AI references, communication efficiency, digital twin construction, multi-mode agreement, opinion dynamics
News Source: Denise Maddox. (November 21, 2026). Targeted human–AI framework cuts construction planning communication in half during simulations. Scienmag.



