AI-Enabled Decision Analytics for Trustworthy Vehicle Digital Twin Governance: A Human-in-the-Loop Framework for Firm-Level AI Absorption, Calibration, and Model Diagnosis

Abstract

European artificial-intelligence policy increasingly combines trustworthy AI, capability building, digital sovereignty, and economic security, yet the availability of regulation, infrastructure, and advanced tools does not guarantee that firms can integrate and govern them effectively. This provision-absorption gap is especially consequential in industrial settings where AI-supported decisions act on proprietary cyber-physical assets. This study develops an AI-enabled decision analytics framework in which a residual-guided Local Artificial Intelligence (AI) Agent supports human-in-the-loop governance of Vehicle Digital Twin (VDT) calibration and model diagnosis. Local execution is used as a controllable data and intellectual-property boundary, while deterministic engineering computation is separated from agentic reasoning so that numerical outputs remain independently auditable. The quantitative proof-of-concept uses a planar vehicle model with lateral and yaw dynamics, a first-order steering actuator, and a known 120-ms steering delay. Thirty Monte Carlo noise realizations are evaluated using calibration and holdout manoeuvres. When the model structure is correct, bounded calibration recovers front and rear cornering stiffness and steering time constant close to their reference values (80.04±0.45 kN/rad, 90.06±0.74 kN/rad, and 0.1002±0.0008 s). At 15 m/s, holdout yaw-rate Root Mean Square Error (RMSE) decreases from 0.0410 to 0.0030 rad/s and lateral-acceleration RMSE from 0.516 to 0.050 m/s². When the steering delay is omitted, optimization still reduces error but creates compensating parameters, including a 96.9% increase in steering time constant. The benchmark therefore demonstrates a firm-level governance mechanism for distinguishing trustworthy model updates from misleading calibration success without delegating engineering authority to the AI Agent.