Skip to content

Content Library

Back to Articles

Article September 9, 2026

Why AI Model Risk Needs a New Rulebook

Why AI Model Risk Needs a New Rulebook

Financial institutions are struggling to fit rapidly evolving AI into traditional model risk frameworks. Risk leaders warn that interconnected agents, dynamic data, third-party models and continuous change require a shift toward system-level governance, automated monitoring and proportionate oversight without sacrificing human judgment.



  • AI should increasingly be governed as an interconnected system rather than isolated models
  • Common data and technologies could create hidden dependencies between AI agents
  • Dynamic data governance and output monitoring are becoming essential
  • Institutions must balance AI innovation against uncertainty and model risk
  • Vendor AI requires controlled testing before access to sensitive data or systems
  • Traditional periodic monitoring may be too slow for rapidly evolving AI
  • Automated systems may increasingly monitor other AI systems
  • Human expertise remains essential but must be deployed selectively

Financial institutions need to rethink model risk management around AI as an interconnected system rather than simply forcing increasingly sophisticated technology into frameworks designed to govern individual models.

As agents proliferate and share common data, tools and methodologies, risks that appear separate could prove dangerously dependent on one another.

Speaking at CeFPro’s 2026 Risk Americas event, an SVP and Head of Line of Business Model Risk Management at a major U.S. bank warned that institutions may be underestimating those dependencies.

Multiple AI agents deriving decisions from similar concepts, tools and data cannot necessarily be treated as unrelated exposures.

“We have N agents, and we have N times X amount of risk we’re carrying. I don’t think that’s quite right,” the senior model risk executive said.

The comparison carries uncomfortable historical echoes. Drawing on his experience during the global financial crisis, he recalled how assumptions around correlation ultimately proved problematic.

AI could create another environment in which individually manageable exposures become considerably more dangerous when their underlying dependencies emerge.

A Head of Machine Learning and Model Governance at a financial services company argued that generative AI also requires dynamic data governance.

Institutions may not fully understand the internal workings of black-box systems, making strong controls around inputs and monitoring of outputs increasingly important.

A Director of Controls at a financial risk advisory firm similarly warned against excessive dependence on individual large language models, particularly for customer-facing applications.

Benchmarking alternatives and avoiding concentration around one model should form part of effective model risk management.

More fundamentally, she argued, “AI is a systems problem, not just a model problem.”

Traditional MRM typically examines individual models deeply, while AI demands greater attention to use cases, interdependencies and the different models operating within an interconnected environment.

That creates a fundamental problem of speed. AI tools can emerge and improve within weeks while established validation processes may take months.

The bank’s senior model risk executive said institutions are still learning how to reconcile that mismatch while ensuring new technology satisfies established compliance obligations.

“We’re learning,” he said. “We’re not there yet.”

The challenge is to prevent governance becoming either an obstacle to useful innovation or incapable of controlling it.

Panelists repeatedly returned to risk and reward, arguing that institutions should determine whether AI generates sufficient value to justify the additional uncertainty it introduces.

The Director of Controls characterized AI as an opportunity for model risk teams rather than simply a threat.

AI could accelerate documentation and generate alternative models capable of challenging existing approaches, allowing specialists to devote more attention to problems previously crowded out by administrative work.

“Years of experience cannot be replaced with just two lines of prompts,” she cautioned, but alternative models and approaches could provide valuable challenges to established thinking.

Oversight must evolve alongside those capabilities. Customer complaints should not become the mechanism through which institutions discover that an AI application is failing.

Internal testing, scenario analysis and continuous monitoring should identify problems before systems reach customers.

Events you may be interested in

View all of our upcoming events

Sign Up to Our Newsletter

Sign up

Connect Plus

Unlock more with Connect Plus membership

This content is available to Connect Plus members.