The Risk Manager Ensuring AI Doesn't Create the Risks It Claims to Manage

By Daniel Foster, FRM, CFA — 2025-12-25

Daniel Foster has spent twenty years in risk management — market risk at a global investment bank through the 2008 crisis, then enterprise risk at a leading fixed income manager. He has watched sophisticated quantitative models fail in ways their builders didn't anticipate more times than he can count. He joined IXO with a specific mission: to ensure that AI risk tools are evaluated by someone who has lived through the tail events the models aren't built to survive.

The epistemological problem

"Risk management AI has a fundamental epistemological problem," he says. "It's trained on historical data, which means it's excellent at managing the risks that have already happened and poor at managing the risks that haven't happened yet. A model trained on post-2008 data knows what a mortgage crisis looks like. It doesn't know what a central bank digital currency crisis looks like, or a sovereign AI arms race crisis, or any of the tail events that are genuinely novel."

Foster's IXO work focuses on risk model evaluation — testing AI risk assessment tools for the quality of their scenario analysis, the appropriateness of their correlation assumptions, and critically, whether they communicate model uncertainty in ways that support rather than undermine good risk decisions.

The VaR problem persists

"The VaR problem is still with us. Value at Risk is a confidence interval, not a worst-case scenario. AI risk tools that present VaR numbers without contextualizing what the distribution looks like beyond that confidence interval are creating false precision. Annotating that gap requires someone who understands why it matters."

Model governance

He also evaluates AI tools in the model validation context — specifically testing whether AI model risk frameworks correctly identify when a model's underlying assumptions have been violated and escalation is required. "Model governance is the unsexy part of risk management that prevents disasters. AI that gives model risk committees false confidence is dangerous."

Read The Risk Manager Ensuring AI Doesn't Create the Risks It Claims to Manage on the IXO blog