Model risk management in the new era of non-traditional AI

What it is about

Walter Paiva Ramos, from Banco de Crédito del Perú (BCP), presented at the Fourth Annual Risk Conference (September 17, 2026) how model risk management changes when artificial intelligence stops being a predictive model and becomes a system. He clarifies that the opinions are personal and academic and do not commit BCP.

The 9 slides propose distinguishing traditional AI, centered on the model, from non-traditional AI, centered on the system, with generative and agentic AI. They review the evolution of AI and of model risk management regulation, from the Federal Reserve SR 11-7 to the 2026 update, and propose definitions of model, AI system and model risk, and an integrated AI governance in which model risk management is one piece.

Objectives

• Distinguish traditional AI, centered on the model, from non-traditional AI, centered on the system.
• Place model risk management in the global regulatory evolution.
• Propose definitions and an integrated AI governance scheme with coordinated roles.

Key takeaways

1. In traditional AI the model is the dominant risk object (logistic regression, trees, GBM, neural networks); in non-traditional AI the model is only one component of a system with data, prompts, memory, rules, APIs and users.
2. In traditional AI, models were inventoried; in non-traditional AI it is better to inventory the use case or system first and then map its models and controls.
3. Model risk management steps in when the AI system carries material model risk, and the risk of the system can exceed that of the model.
4. The new challenge is moving from governing models to governing systems: more complexity, more actors and more coordination with non-financial risk functions.
5. The presentation draws on a working paper by Paiva and Breeden (2026) on integrated AI governance, which moves from isolated controls to integrated governance.