Skip to content

Content Library

Back to Articles

Article January 13, 2026

Governing Generative AI Without Breaking Model Risk Foundations

Governing Generative AI Without Breaking Model Risk Foundations

As generative AI moves from hype to deployment, large banks are extending mature model risk frameworks rather than rebuilding them. A senior model risk leader explains why governance, adaptability, and human judgment are becoming more important as institu
  • Generative AI has moved from experimentation to disciplined deployment
  • Banks are extending existing model risk frameworks rather than rebuilding them
  • Validation teams play a central role in shaping Gen AI governance
  • Regulatory expectations are evolving and require adaptable policies
  • Industry approaches to Gen AI risk are converging across large banks
  • Monitoring frameworks are being extended to capture new failure modes
  • Evaluating Gen AI outputs remains complex and human intensive
  • Automation of evaluation is necessary but introduces new risks
  • Gen AI currently increases oversight demands rather than reducing them

The surge of enthusiasm around generative AI over the past year has been dramatic, moving quickly from experimentation to structured implementation across financial services.

For risk leaders inside large banks, the challenge has not been whether to engage with Gen AI, but how to govern it responsibly without destabilizing long established control frameworks.

Speaking at a New York conference of senior risk managers, the director of model validation at a major US bank described the current phase as a necessary shift from excitement to discipline.

“I speak from experience when I say we’ve moved past the initial euphoria and into a period where governance, controls, and accountability really matter,” he told delegates.

Rather than treating generative AI as a wholly new risk category, his team has focused on extending existing model risk management foundations.

Years of experience validating traditional models have provided a strong base for addressing Gen AI risks such as bias, fairness, explainability, and toxic outputs.

The executive emphasized that reinventing frameworks would create unnecessary fragmentation. The priority instead has been to layer new controls onto proven structures.

Model validation sits at the center of this effort. While governance teams own formal policies and procedures, validators play a critical role in shaping how new risks are identified and operationalized.

According to the speaker, this work goes far beyond technical testing. It also involves influencing policy design, collaborating across functions, and embedding controls that reflect how models are actually used in business contexts.

That collaborative approach, he says, has become more important as regulatory expectations evolve. Supervisory guidance on artificial intelligence is still emerging, and standards differ across jurisdictions.

The speaker noted that adaptability is essential. “We cannot afford to be reactive,” he said. “We design our policies knowing that regulation will change, and we need to absorb that change smoothly.”

Large institutions have an advantage in this environment because they can draw on expertise from legal, compliance, technology, and risk teams simultaneously.

When new guidance appears, it can be interpreted and integrated quickly rather than bolted on as an afterthought. This, he argues, helps to maintain consistency across the model lifecycle and avoids creating parallel governance tracks for Gen AI.

Industry dialogue has also reinforced a sense of convergence. Through discussions with peers at major regulatory and industry events, the executive told delagtes he had observed that large banks are approaching Gen AI with similar levels of caution.

While terminology and internal mechanics may differ, the underlying philosophies are closely aligned, and that consistency, he argued, is a positive signal for supervisors and the broader market.

Ongoing monitoring remains anchored in established supervisory principles. Frameworks developed under long standing guidance continue to perform well, but they must now account for what makes Gen AI different.

New risks emerge in production, new failure modes appear, and performance indicators are more complex.

Still, the answer is extension rather than replacement, the speaker said. Robust monitoring systems already exist and can be adapted to reflect these new realities.

The most difficult challenge ahead lies in evaluation. Traditional models can often be assessed using familiar statistical metrics.

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.