AI Governance
Boards Are Approving AI Strategies They Cannot Interrogate
The gap between AI deployment and AI governance in financial services is widening faster than most risk functions can absorb. Boards are signing off on AI strategies presented in the language of competitive advantage, without the vocabulary or the frameworks to interrogate what they are actually approving.
This is not a technology problem. It is a governance problem — and in a regulated financial services environment, governance problems have capital consequences.
What Board-Level AI Accountability Actually Requires
A clear answer to the question: if this model causes harm, who is accountable? Not which team owns the model — which named individual carries the fiduciary responsibility for its outcomes.
Model risk taxonomy that maps AI risk categories to existing risk frameworks — not a separate AI risk register that sits outside the second line of defence.
Board reporting that translates model behaviour into fiduciary language. Accuracy rates and F1 scores are not board metrics. Cost of model error, regulatory exposure, and customer impact are.
An escalation path for algorithmic decisions that cause unexpected outcomes. This path needs to exist before the outcome occurs.
A definition of explainability that would satisfy your regulator, not just your data science team.
Why Most Current Frameworks Will Not Survive the First Regulatory Challenge
Most AI governance frameworks in South African financial services have been built by technology teams for technology teams. They govern model development and deployment well. They govern accountability, explainability, and board-level oversight poorly — because those disciplines require governance expertise that technology teams do not typically carry.
The first FSCA or SARB action against an institution for AI-related harm will expose this gap very quickly. The institutions that will weather that moment are the ones that designed their AI governance at the board level first, and delegated the technical implementation downward — not the other way around.