The explainability overhead
For high-risk decisions, governance mandates that every prediction be explainable. While the libraries used to achieve this (like the popular SHAP and LIME) are open source, they are not free to run. They are computationally intensive. In practice, this means running a second, heavy algorithm alongside your main model for every single transaction. This can easily double the compute resources and latency, creating a significant and recurring governance overhead on every prediction.
The continuous monitoring burden
Standard MLOps involves monitoring for performance drift (e.g., is the model getting less accurate?). But AI governance adds a second, more complex layer: governance monitoring. This means constantly checking for bias drift (e.g., is the model becoming unfair to a specific group over time?) and explainability drift. This requires a separate, always-on infrastructure that ingests production data, runs statistical tests and stores results, adding a continuous and independent cost stream to the project.
The audit and storage bill
To be auditable, you must log everything. In finance, regulations from bodies like FINRA require member firms to adhere to SEC rules for electronic recordkeeping, which can mandate retention for at least six years in a non-erasable format. This means every prediction, input and model version creates a data artifact that incurs a storage cost, a cost that grows every single day for years.
