Published Paper Details:

Advanced AI systems used in financial services increasingly rely on extended, autonomous reasoning rather than simple pattern recognition or prediction. These reasoning models generate intermediate deliberations--commonly referred to as chains of thought (CoT)--before producing outputs or taking actions. Recent research suggests that monitoring these reasoning traces can significantly improve the detection of misbehavior, misalignment, or unintended objectives compared to traditional output-only monitoring approaches [P1], [P3]. For model risk management (MRM), validation, and AI governance professionals, this presents both an opportunity and a risk. CoT monitoring provides a partial window into how advanced models reason, enabling earlier identification of behaviors such as reward hacking, deception, or policy-violating intent that may not be observable in final outputs alone [P1], [P3]. At the same time, the ability to monitor reasoning is not guaranteed to persist. Research shows that chains of thought are often incomplete, selectively unfaithful, or absent in precisely those settings where reliable oversight is most needed [P5]. Training choices, architectural shifts, and optimization pressures can further degrade monitorability over time [P1], [P4]. This white paper translates the emerging CoT monitorability literature into a governance-ready framework for financial institutions. It clarifies what CoT monitoring is--and is not--capable of delivering; outlines where it adds the most value for high-risk banking use cases; and proposes how monitorability can be operationalized as a distinct risk attribute within existing enterprise AI governance and model risk programs. Throughout, the paper emphasizes that CoT monitoring should be treated as one layer in a defense-in-depth strategy, rather than as a substitute for validation, testing, or explainability controls