Skip to content

Content Library

Back to Articles

Article January 23, 2026

Artificial Intelligence Upends Bank Risk Governance

Artificial Intelligence Upends Bank Risk Governance

As artificial intelligence reshapes financial decision making, traditional model risk frameworks are being stretched to breaking point. At a recent flagship US conference, senior risk leaders debated how banks should govern AI tools whose scale, opacity,
  • AI tools are straining traditional model risk frameworks built for statistical models
  • Definitions of model versus tool are no longer sufficient for modern analytics
  • Data quality and governance remain the primary source of AI risk
  • Vendor provided models introduce opacity and intellectual property concerns
  • Risk ratings must reflect use case impact rather than penalizing AI uniformly
  • Monitoring needs to be more frequent and use case specific
  • Collaboration across lines of defense is essential for effective governance
  • A unified analytical risk framework is emerging as the long term solution

Senior risk managers have been warned that artificial intelligence is forcing banks to rethink how model risk, governance, and accountability are defined.

At a CeFPro conference on vendor risk, one session explained how, as machine learning moves from experimentation to embedded decision making, long-standing validation frameworks are being stretched, exposing gaps in oversight, explainability, and control that institutions can no longer afford to ignore.

The discussion, led by senior heads of modeling and model risk at large US and international banks, made clear that many firms are trying to stretch frameworks designed for statistical credit and market models to fit tools built on machine learning, large language models, and vendor provided analytics.

The result, participants suggested, is growing friction between innovation and control.

A senior modeling leader at a mid sized US consumer finance institution described how her organization manages more than one hundred models across regulatory, pricing, and marketing use cases.

From her perspective, the first line can no longer treat AI governance as something handed off downstream.

Controls, documentation, and risk mitigation must be embedded at design stage, even when regulations remain fragmented across states and use cases.

The emphasis, she argued, is less about regulatory minimums and more about doing the right thing for customers and employees.

Much of the tension centers on definitions. Several speakers noted that internal debates over what constitutes a model versus a tool have returned with greater force in the AI era.

Not every statistical technique is artificial intelligence, and not every AI enabled application should automatically trigger the highest risk rating.

Without a shared language, firms struggle to conduct meaningful gap and risk assessments, particularly when tools are sourced from vendors and embedded across multiple workflows.

A head of model risk management at the US branch of a large international bank traced this problem back to regulatory history.

Model definitions established more than a decade ago still underpin supervisory expectations, yet the population of analytical tools has expanded dramatically.

End user computing tools, vendor hosted AI services, and generative applications now sit uneasily alongside traditional valuation and credit models. According to the speaker, the boundary between model and non model is no longer a helpful organizing principle.

Data emerged as the most consistent point of concern. Participants repeatedly emphasized that no AI control framework can compensate for weak data governance.

Training data quality, representativeness, bias, noise, and intellectual property rights all introduce material risk, especially when firms rely on vendor models trained on opaque datasets.

Several speakers warned that fine tuning external models with poor internal data can compound rather than mitigate risk.

Usage context also matters. Internal productivity tools such as copilots present a different risk profile from customer facing applications like chatbots or automated decision engines.

While internal tools can still create harm through inaccurate reporting or poor quality outputs, external tools raise immediate legal, reputational, and conduct risks.

Risk ratings, speakers argued, must reflect impact and frequency rather than penalizing all AI techniques equally.

Monitoring expectations are also shifting. Traditional models may have been reviewed periodically, but AI systems evolve continuously. Speakers highlighted the need for more frequent testing of robustness and accuracy, along with clear metrics aligned to each use case.

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.