Digital twins for operational risk: regime-aware routing in financial workflows

Human-centric financial services workflows exhibit regime-switching: a ‘normal’ mode of routine cases and an ‘exception’ mode with heavy tails and volatility. Digital twins fitting only average regimes can reproduce central tendencies while systematically underpredicting congestion and risk. We construct a closed-loop stochastic organisational digital twin integrating (i) process-informed structure, (ii) bimodal timing inference via a Gaussian mixture model (GMM) on log-transformed service times, and (iii) bidirectional control through Regime-Aware Dynamic Routeing (RADR), which isolates exception-prone work. To ground the structural assumption of bimodality, we evaluate the real-world Business Process Intelligence (BPI) 2017 event log, where a two-component GMM reduces Kolmogorov-Smirnov (KS) error by 56.55% (from 0.1347 to 0.0585) compared to the unimodal baseline. Evaluating the RADR control policy within a calibrated, synthetic environment (training n=70000, test n=30000; seed 42), the GMM yields a held-out KS statistic of 0.0038, compared with 0.1498 for the unimodal baseline (97.5% reduction). In queueing simulations, RADR reduces peak underwriting backlog from 42,207 to 12,526 cases and peak work-in-process from 42,212 to 12,532, with an average 0.378/hr reroute rate. Results show regime-aware stochastic modelling and control are necessary to obtain reliable organisational digital twins in workflows where exceptions drive tail risk.