The Irrecoverable Institution -  Why Replayability - Not Explainability - Is the Governance Standard for Agentic Banking 

Abstract

Financial institutions have accelerated the deployment of Agentic AI systems. These systems have the capabilities to make decisions and are operating faster than any governance mechanism can respond. However, when such decisions go wrong, the biggest cost incurred is not due to their very first mistake, but because the financial institution lacks the capability of replaying and governing the decision path until other costs increase significantly.

The objective of this paper is to propose Decision Replay as the base governance framework for agentic financial institutions, which would be the capability to replay and govern decisions at any moment, via any number of AI agents, back to the policies and authorization behind such a decision. While the explainability framework (LIME, SHAP, and interpretable model) provides information as to how and why a certain model came up with that decision, the question about whether it was a properly governed, trackable, and authorized process cannot be answered with explainability frameworks.

Four different case studies of failures are explored in this paper: Silicon Valley Bank (March 2023),  Knight Capital Group(August 2012), the "Gilt" market collapse in the UK (September 2022), and the credit agentic decision failure.

In each of the four failures studied above, 35% to 60% of the firm's losses are a direct result of failing to track, audit, and act on decisions.

EU AI Act, DORA, MAS FEAT, and Basel guidance are examined.

The paper builds on previous work on the Decision Integrity Chain™ (DIC™)- an eight-layer governance framework and introduces the FUSE/STAGE architectural stack that operationalizes it. 

Keywords: Decision Replay, Agentic AI governance, Replayability, Decision Integrity Chain™, AI accountability, Institutional governance, Model risk, Explainability JEL Classification: G21, G28, G32, L86, K22

JEL Classification: G21, G28, G32, L86, K22