Stable Decision Routing in Agentic Financial AI via Self-Reported Uncertainty Signals

Deploying autonomous AI agents for financial decisions—such as credit assessment—requires reliable mechanisms to route uncertain cases to human review. We propose a lightweight approach that uses LLM self-reported uncertainty as a proxy control signal for decision routing, without requiring access to model internals or ground-truth labels. Through experiments on 300 synthetic credit applications with four perturbation types (weak to adversarial), we demonstrate that LLM-based signals achieve 69% lower flip rates than heuristic baselines at comparable review rates (τ =0.5), with medium effect size (Cohen’s d=0.507). Our results show that self-reported uncertainty enables stable, controllable routing policies that degrade gracefully under distribution shift—a key requirement for responsible deployment of financial AI agents. We emphasize that our goal is not uncertainty calibration, but controllable human–AI handoff in agentic systems.