A hybrid AI model flags construction settlement disputes two weeks earlier and cuts auditor time by about two-thirds in a 150-node case study, achieving 92.5% accuracy; the promising results rely on a small, domain-specific sample and require broader validation before industry-wide claims.

Main Finding

A machine-learning numerical–semantic dual-path model (XGBoost for numeric features + Attention-BiLSTM for text features, fused via a Gated Fusion Unit and tuned with Bayesian optimization) can accurately and early identify cost-dispute risks in construction settlement auditing. On a 150-node real-world dataset it achieved F1 = 0.901 and accuracy = 92.5%, reduced per-audit time by 65%, and provided warnings on average 14 days earlier than existing practice — substantially outperforming single-path approaches.

Key Points

• Problem: Late and error-prone detection of cost-dispute risks in construction project settlement audits due to heterogeneous multi-source data (numbers + contract/change texts + environment).
• Feature engineering: High-dimensional vectors built from three dimensions:
  • Numerical sensitivity (quantity and price deviations),
  • Semantic conflict (textual fingerprints of rights/responsibilities conflicts),
  • Project environmental background.
• Model architecture:
  • Path A: XGBoost to capture explicit numerical risk signals.
  • Path B: Attention-BiLSTM to extract implicit semantic conflict signals from contract and change-order texts.
  • Fusion: Gated Fusion Unit (GFU) to non-linearly combine numerical probabilities and semantic indices, enabling trade-offs and joint representation.
  • Optimization: Bayesian optimization for global hyperparameter tuning.
• Evaluation:
  • Dataset: 150 real settlement nodes (case study).
  • Metrics: F1 = 0.901, Accuracy = 92.5%.
  • Practical impact: 65% reduction in auditor time per case; average 14-day earlier warnings.
• Comparative result: Clear improvement over traditional single-path (numeric-only or text-only) identification models.

Data & Methods

• Data sources: Engineering quantity and price records, contract texts, change order texts, and contextual project/environmental metadata (three-dimension inputs).
• Feature extraction:
  • Numerical: sensitivity measures capturing deviations in quantities and prices.
  • Textual: semantic conflict indicators via sequence modeling and attention mechanisms.
  • Environmental: contextual features affecting dispute likelihood.
• Modeling:
  • Gradient-boosted trees (XGBoost) handle structured numeric inputs and yield probability/risk scores.
  • Attention-enhanced BiLSTM processes sequential contract/change-order text to produce semantic conflict indices.
  • GFU performs gated, non-linear fusion of outputs from both paths into a combined risk prediction.
• Training & tuning: End-to-end/global parameter tuning with Bayesian optimization to balance performance and avoid overfitting.
• Evaluation approach: Case-study testing on 150 settlement nodes with standard classification metrics (precision/recall/F1/accuracy) and operational measures (audit time reduction, advance warning days).
• Baseline: Traditional single-path identification models (numeric-only or semantic-only).

Implications for AI Economics

• Productivity & cost savings:
  • Large per-audit time reductions (reported 65%) imply significant labor-cost savings and higher throughput for audit firms and owners.
  • Earlier warnings (≈14 days) allow preemptive mitigation, potentially lowering dispute resolution costs and project delays.
• Market structure & incentives:
  • Increased automation could shift auditor labor demand toward higher-skill roles (model validation, exception handling), compressing demand for routine audit tasks.
  • Audit firms that adopt such tools may gain competitive advantage, prompting industry-wide adoption and new service offerings (automated pre-audit, continuous monitoring).
• Risk allocation & contracting:
  • Better early detection changes bargaining dynamics and may reduce incidence/severity of disputes; contracting parties may update clauses, pricing strategies, or insurance terms accordingly.
• Returns to data & complementarities:
  • Value of rich multi-source data is highlighted — firms with better digital records capture more benefit, creating incentives to invest in standardized, machine-readable recordkeeping.
  • Complementarity between numeric and textual data shows higher returns to integrated analysis than to isolated analytics.
• Policy, governance & standards:
  • Widespread deployment would benefit from standards for data formats, labeling, and model auditing to ensure fairness, transparency, and robustness.
  • Regulatory acceptance (e.g., of automated early warnings) may be needed for formal audit workflows.
• Risks & adoption barriers:
  • Small study size (150 nodes) and domain specificity imply need for larger, cross-project validation before broad economic conclusions.
  • Potential model biases or misclassification could create new risks (false positives/negatives), requiring human-in-the-loop workflows.
  • Data privacy, proprietary contract content, and integration costs could slow adoption.
• Research & evaluation priorities for AI economics:
  • Quantify macro adoption effects: industry-level labor shifts, price effects for audit services, dispute incidence changes.
  • Cost-benefit and ROI analyses across diverse project types and jurisdictions.
  • External validation on larger and more varied datasets; study generalizability across contract forms and procurement regimes.
  • Investigation of incentive effects (e.g., strategic reporting by contractors) and mitigation (auditability, explainability).

Limitations to note: promising case-study results but limited sample size; model complexity (GFU + neural text path) may reduce interpretability relative to simpler rule-based audits; operational integration and robust evaluation in live environments remain necessary before large-scale economic claims.