AI’s Emerging Frontier in Balance Sheet Risk Management: Opportunities and Challenges

In short order, AI has risen to the top-tier spot on every boardroom’s emerging-risk dashboard, and few would argue its presence. AI is one of the few forces in banking that carry such perfect balance on the classic risk-return trade-off scale, opportunities on one side and dangers on the other. How banks strike that balance may separate those that pull ahead from those that merely keep pace.

AI conversations often include a suite of terms such as machine learning, deep learning, generative AI, data science, large-language models, and so on used interchangeably, which may lead to confusion when communicating to executive audience. To navigate this, I think of AI into two broad families: Predictive AI and Generative AI. Predictive AI forecasts the future from historical data by identifying patterns and trends. Generative AI creates new content such as text, images, video, audio, or code. Keeping clear distinction removes buzz-word distractions and frames the practical questions addressed here.

Real and immediate upsides

Right now, the most tangible gains lie on the predictive side, because banks have relied on statistical models for decades. What has really changed is the speed and reach of modern algorithms, which can now solve sophisticated problems for institutions of all sizes, capabilities once limited to banks that invested heavily in this area. Large-scale, complex balance-sheet optimizations under a suite of capital, liquidity, interest rate risk, and funding constraints are now achievable with deep learning, a Predictive AI technique. Predictive AI models are also becoming prevalent in mortgage prepayments to deposit behavior modeling. However, Generative-AI use is still narrow such as for text classification to track trader or customer patterns or summarize dense documents. Lack of fully effective control frameworks for Gen. AI slow adoption for decision making beyond these use cases.

Guard-rails and Transparency to ensure safety and speed

The Predictive-versus-Generative split is useful for governance. Banks have frameworks for Predictive AI model governance since SR 11-7 was issued more than a decade ago; that rulebook for model risk management remains the field’s original responsible AI manifesto written long before responsible and ethical AI buzzwords were coined. Most of that governance framework works well for predictive AI and with a few extra bolts and better plumbing, it can be modernized for Generative AI’s governance. Generative AI needs to be managed with the same impact-based tiering banks already use for their financial models. The impact should be carefully assessed here because the impact of Generative AI goes beyond financial risks or decisions - into lawsuits, headlines, reputation, and public trust. Once a tiering structure is in place, high-impact models or tools should merit deep validation like any other models while exploratory idea models, that are in a tightly bound sandbox, can continue to experiment with light touch reviews. Institutions should be careful of calling some of these AI tools as non-models and skip validations and model risk governance – that is a form of model definition arbitrage which could expose banks to risks they do not intend to take. There are valid reasons some banks do this as burden of model governance can stifle innovation. But this is the job of model risk managers to cultivate a culture of risk-based monitoring and use approaches that are commensurate with impact and allow for value creation.