Traditional model risk management frameworks are struggling to accommodate rapidly evolving AI technologies. A senior banking executive argues that financial institutions must replace static validation with continuous monitoring, strengthen cross-functional governance and prioritize high-risk models. The discussion highlights why transparency, business accountability and new leadership approaches are essential to effective AI oversight.
Banks must fundamentally rethink model risk management as
artificial intelligence challenges the assumptions underpinning traditional
validation, governance and oversight.
Rather than attempting to impose established controls on
fundamentally different technologies, risk leaders need to develop new
approaches built around continuous monitoring, collaboration and a clearer
understanding of emerging threats.
Speaking at CeFPro’s 2026 Risk Americas event, the managing
director and global head of model risk management at a major international bank
argued that financial institutions risk applying outdated methodologies to
technologies whose decision-making processes may be impossible to explain using
conventional techniques.
"When you get to the question of validating the agents
or validating AI tool, it's more of a question, almost like a philosophical
question," he said.
Traditional model validation assumes that risk managers can
examine a model's underlying logic, assess its conceptual soundness and
determine whether its outputs are reliable. Advanced AI systems challenge that
assumption.
The executive suggested that understanding precisely how an
AI agent reaches a decision may be beyond the capabilities of conventional
validation.
"You will never be able to really validate it in a
common sense of the word, like we use for traditional models," he warned.
From Reactive Oversight to Proactive Leadership
The discussion highlighted a fundamental weakness in
established model risk management practices: institutions frequently react to
technological developments rather than anticipate the governance challenges
they create.
Although risk leaders aspire to proactive oversight, the
executive acknowledged that much of the industry's response remains reactive.
"We want to be proactive, but in reality, we are
reactive," he said.
Moving beyond that position requires more than introducing
additional controls. Risk professionals must acquire new skills, challenge
established assumptions and formulate different questions before attempting to
develop solutions.
The executive compared the emergence of AI with a fundamental
change in scientific understanding, arguing that conventional analytical
frameworks may be inappropriate for technologies operating according to
unfamiliar principles.
This creates a significant leadership challenge.
Senior management must understand why existing frameworks are
insufficient without becoming overwhelmed by technical explanations.
The executive emphasized the importance of communicating
these challenges in accessible terms to secure investment, specialist expertise
and organizational support.
AI Governance Cannot Remain in Silos
Accountability becomes increasingly complicated as AI
applications spread across model risk, technology, data management and
operational risk functions.
The executive rejected the idea that institutions should
divide responsibility according to traditional departmental boundaries.
"It's obviously, it should not be siloing," he
said, emphasizing the need for collaboration between functions possessing
different forms of expertise.
Recent changes in model risk guidance have intensified
questions about responsibility for generative and agentic AI.
The executive noted that excluding certain AI technologies
from traditional model risk requirements does not eliminate their risks.
Instead, institutions must determine which governance arrangements are
appropriate.
This is particularly challenging for multinational banks
operating across jurisdictions with different regulatory expectations.
He argued that AI governance ultimately requires an
interdisciplinary approach, potentially supported by dedicated councils or
forums bringing together representatives from multiple risk functions.
Such structures could help institutions identify emerging
applications, determine appropriate ownership and coordinate oversight.
Continuous Monitoring Must Replace Static Validation
Perhaps the most significant operational change concerns the
frequency and purpose of model validation.
Traditional approaches rely heavily on periodic reviews,
often conducted annually or over longer intervals.
However, AI systems may change too quickly for these
processes to remain effective.
"All this kind of periodic validation becomes really
nonsensical for that sort of tools," the executive said.
Instead, institutions should place greater emphasis on
continuous performance monitoring and the assessment of model outputs.
This would shift attention away from repeatedly examining
conceptual soundness toward determining whether systems continue to produce
appropriate results under changing conditions.
The executive acknowledged that monitoring presents its own
difficulties, particularly when establishing meaningful benchmarks for
AI-generated outputs.
Nevertheless, he argued that regular observation of model
performance is becoming essential.
Even models used infrequently may become unreliable between
executions as market conditions and operating environments change.
The challenge is therefore to develop controls proportionate
to each model's characteristics, materiality and potential consequences of
failure.
Transparency Is the Foundation of Effective Oversight