What is AI model risk management?
By Nick Jang, founder of MINERVA Studio ·
The short answer
AI model risk management is model risk management extended to machine learning, generative AI and AI agents: the same inventory, rating, independent review, approval and monitoring, plus controls for what AI adds, namely opacity, drift, bias, non-deterministic outputs, dependence on vendors’ foundation models and, for agents, the ability to act. In Canada, OSFI’s E-23 brings AI inside model risk from 1 May 2027.
Why AI changes model risk
A logistic regression scorecard has a dozen inputs and a validator can read every coefficient. A gradient-boosted model might have three hundred features interacting in ways nobody can write down. A large language model has billions of parameters trained on data your bank has never seen. The controls are the same kind of controls. They just get much harder to apply.
| Property | Why it matters | What to add |
|---|---|---|
| Opacity | Hard to explain why the model produced a result, to a customer or a validator. | Explainability methods (such as SHAP), documented limits, human review for high-impact decisions. |
| Drift | Learned patterns decay as behaviour changes; some models retrain on their own. | Tighter monitoring of inputs and outputs, retraining triggers, review of each retrained version. |
| Bias | Models can reproduce unfair patterns in historical data. | Fairness testing across groups, data controls, a record of the trade-offs chosen. |
| Non-deterministic outputs | A generative model can answer the same question differently, or invent facts. | Evaluation sets, hallucination and accuracy measures, output filters, grounding in approved sources. |
| Vendor foundation models | The model may belong to a provider who can change it without notice. | Vendor model treated as a third-party model under E-23 and B-10, with change notification in the contract. |
| Autonomy | An agent acts through tools: it sends, pays, files, changes records. | Policies on what each agent may do, human approval for high-risk actions, a kill switch, logs someone reads. |
What OSFI E-23 says about AI
E-23 puts AI and machine-learning methods inside its definition of a model, so every rule applies to them. It also asks for specific things where AI is concerned: when a model owner states a model’s purpose and risk, that includes transparency, explainability, bias, ethics and privacy; data standards cover bias and controls over synthetic and proxy data; documentation covers explainability; monitoring covers AI-specific risks such as drift; and there are extra controls where a model is a black box or acts on its own.
For the agents themselves, OSFI’s July 2026 bulletin on generative and agentic AI adds technology and operational expectations: each agent with its own identity, least-privilege access, an allow-list of tools, approval for high-impact actions and reviewed logs. Our guide to agentic AI governance covers that half.
Adapting each step of MRM for AI
- Inventory. List AI models with their provider, version and the systems and agents that use them. Generative AI tools adopted by teams belong on it, not just models built by data science.
- Risk rating. Add factors for autonomy (does it decide or act without a person?), customer impact, explainability and reliance on a vendor.
- Development standards. Require evaluation sets, documented prompts and configuration for generative systems, and training data lineage where the institution controls it.
- Validation. Test performance, robustness, fairness and, for generative AI, accuracy against ground truth, hallucination, prompt injection and leakage. See model validation.
- Approval. Approve the use, not just the model: a model approved for internal summaries isn’t approved for advice to customers.
- Monitoring. Watch input drift and output quality continuously, sample outputs for human review, and treat each retrained or vendor-updated version as a change.
The United States: SR 26-2
On 17 April 2026 the Federal Reserve, the OCC and the FDIC replaced SR 11-7 with SR 26-2, revised interagency guidance on model risk management, and withdrew SR 21-8 on BSA/AML models. Commentators describe it as more risk-based and less prescriptive. Their readings of how far it reaches generative and agentic AI differ, so a bank should read the guidance itself on scope. See what SR 11-7 was and what replaced it.
Where MINERVA fits
MINERVA keeps AI models in the same inventory as every other model, with the owner, reviewer and approver separated and enforced, and records validation results from tests run where the data lives, including fairness and LLM suites. AI agents run under written policies, decided by a person, with held calls, a kill switch and a live decision feed. See the agentic AI governance platform.
Related questions
Is AI model risk different from traditional model risk?
It is the same risk, that a model is wrong or misused, with new sources: opacity, drift, bias, unpredictable outputs, vendor dependence and, for agents, actions taken without a person.
Does OSFI E-23 apply to generative AI?
Yes. E-23’s definition of a model includes AI and machine-learning methods, which covers generative models used by a federally regulated institution.
How do you validate a large language model?
With evaluation sets drawn from the intended use: accuracy against known answers, hallucination rate, robustness to adversarial prompts and prompt injection, toxicity and leakage of sensitive data, plus review of how the system grounds its answers.
What is the difference between AI governance and AI model risk management?
AI governance is the broader set of policies, roles and controls for all AI use. AI model risk management is the part that controls the risk of the models themselves. See what AI governance is.
Sources
- OSFI — Guideline E-23, Model Risk Management (2027)
- Federal Reserve — SR 26-2, Revised Guidance on Model Risk Management
- NIST — AI Risk Management Framework: Generative AI Profile (NIST AI 600-1)
New answers are posted on MINERVA on LinkedIn. Follow along, or send the question you’d like answered next to minerva@minerva.studio.
A plain-language explanation, not legal advice. The regulator’s own text is the source; read it, and take advice on how it applies to your institution.