How has the definition of ‘model risk’
expanded since the early 2010s, and what new categories of risk do model risk
teams now have to oversee that simply did not exist when frameworks like SR
11-7 first emerged?
Model Risk has had to adapt along with
the changes in the banking industry. With the adoption of AI and large language
models in business, Model Risk must stay adaptable and intentional to maintain
effective governance frameworks for emerging model types.
With the rapid adoption of AI, machine
learning, and large language models across banking operations, how should model
risk teams rethink traditional validation approaches that were originally
designed for deterministic financial models?
Model risk teams have had to move
beyond the traditional toolbox of statistical testing to include AI specific
risks.
Generative AI models require a deeper
dive into data accuracy and controls, explainability of outputs, and evaluation
of performance that traditional financial models do not require.
Fortunately, for most banks, the
burden is not entirely on the MRM teams – we must work along-side other subject
matter experts in the Bank, like Data Management/Governance and Information
Security/Technology to help ensure the overall governance of AI Models are
effectively managed.
Many banks built their model risk
functions around quantitative finance expertise. As model usage expands into
areas like fraud detection, cybersecurity, and AI-driven decision tools, what
new skills and capabilities are most urgently needed inside model risk teams
today?
The technical and mathematical skills
are in place, and transferrable to evolving models and technologies – at least
for Model Risk Management.
I believe that being adaptable,
curious, and eager to work with others is essential in the bank today.
Establishing proactive working relationships throughout the organization is
essential to ensure that each new model is accurately tracked and effectively
governed.
Validation frameworks and tooling were
largely built for conventional models. In practice, where do existing
validation tools fall short when applied to modern AI systems, and how should
institutions begin to close those gaps?
On the front end, it is important to
have strong intake, review and implementation processes for new AI tools and
technologies. For mid-sized banks, this usually requires an enhancement to IT
and data infrastructures to guarantee proper implementation of these models.
On the back end, user education is a
more difficult challenge. The rapid adoption of these tools throughout what
seems all areas of the bank make the ‘human in the loop’ an essential control
point in AI usage.