Model risk management is undergoing a structural shift as AI adoption accelerates. Odile Hounkpatin explains how governance, validation, and oversight must evolve to address ethical, operational, and societal risks, transforming MRM into a cross-functiona
Model risk management is entering a
period of significant transformation as artificial intelligence becomes
embedded in banking decision-making, expanding both the scope and complexity of
risk oversight.
According to Odile Hounkpatin, Head
of Model Validation at Santander UK, the increasing influence of frameworks
such as the OECD AI Principles is reshaping how institutions approach
governance and control.
“MRM is becoming increasingly
cross-functional,” she explains, noting that legal, procurement, IT, and
compliance teams are now actively involved in oversight alongside traditional
model risk functions.
This shift reflects a broader
reality. AI and generative AI models do not operate in isolation, and their
impact extends far beyond technical outputs into operational, regulatory, and
societal domains.
As a result, institutions are being
forced to rethink how models are catalogued, validated, and monitored.
One of the most important changes
lies in how model inventories are constructed. Firms are no longer simply
tracking models themselves, but also the context in which they are deployed and
the decisions they influence.
This expanded view is essential for
understanding risk exposure across increasingly complex systems.
Validation practices are also
evolving. Traditional approaches, which separated pre-deployment validation
from post-deployment monitoring, are giving way to continuous oversight across
the entire model lifecycle.
“Validation and monitoring practices
are transforming,” Hounkpatin says, emphasizing the need for ongoing assurance
rather than one-time checks.
Despite these changes, the
foundational pillars of model risk management remain intact. Risk
identification, governance, development, validation, and monitoring still form
the core of effective frameworks.
However, these structures must now be
extended to address the unique characteristics of AI, including
non-deterministic outputs and broader ethical considerations.
“An integrated MRM framework should
blend traditional rigour with AI-specific oversight,” Hounkpatin offers,
highlighting the importance of maintaining consistency across both conventional
and AI-driven models.
Central to this evolution is the
challenge of measuring explainability, fairness, bias, and safety. Hounkpatin
points to a combination of quantitative and qualitative approaches as
essential.
Tools such as feature attribution
methods help explain model behavior, while fairness metrics assess whether
outcomes are consistent across different groups.
At the same time, stress testing and
adversarial testing are becoming critical for evaluating robustness under
real-world conditions.
“Effective assessment must combine
quantitative indicators with qualitative evaluation and human oversight,” she
says, underlining the limitations of relying on metrics alone.
Looking ahead, regulatory
expectations are expected to accelerate these changes. What was once guidance
is increasingly becoming formal obligation, driving a shift toward more
structured and comprehensive governance frameworks.
Hounkpatin notes that organizations
are responding by establishing dedicated AI governance committees that bring
together stakeholders from across the business.
These groups are responsible for
ensuring accountability, defining escalation paths, and maintaining
transparency in decision-making.