Evaluating Human–AI Decision Architectures Under Varying Risk and Readiness: A Monte Carlo Simulation in Industrial Quality Control

Artificial Intelligence is increasingly involved in industrial decision-making, but its valuedepends not only on predictive accuracy but on how information and decision power areshared between humans and AI. This study assesses six Human–AI decision architectures forindustrial quality control through a controlled Monte Carlo simulation: Human-only,calibrated cost-aware AI-only automation, Black-box Human–AI, Explainable Human–AI,Actionable Human–AI, and a normative expected-cost benchmark. The analysis combines abaseline experiment, a full-factorial sensitivity analysis over 27 combinations of AIreadiness, human readiness, and false-accept cost, and a 60-month dynamic simulationincorporating human learning, AI performance degradation, and periodic retraining. Nosingle deployable architecture is uniformly best across all conditions and performancecriteria. At baseline, AI-only automation achieves the lowest Cost of Quality amongdeployable architectures, while Actionable Human–AI clearly outperforms Human-only,Black-box, and Explainable designs. However, Actionable Human–AI beats AI-onlyautomation in cost terms in only three of the 27 static scenarios, all involving low AIreadiness with sufficiently capable humans. In dynamic weak-AI conditions, ActionableHuman–AI cuts false-accept rates by over 24 percentage points and improves normativedecision alignment at only modest extra cost. The findings show that explainability alonedoes not improve operational decisions, that human control adds value only under specificcapability and risk conditions, and that AI's role should be treated as a context-dependentdesign factor. The study offers a computational framework for evaluating Human–AIdecision architectures beyond model accuracy alone.