Book a demo
Topics · REGULATORY

Banking Model Risk How It Works | Moring

Banking AI model risk management works as a loop: inventory and risk-tier every model, document it thoroughly enough to be recreated by a third party, validate and approve before deployment, then monitor and revalidate for drift. Generative frameworks extend SR 26-2 concepts into layered controls covering approved use, risk tiers, and evidence.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

Banking AI model risk management works as a loop: inventory and risk-tier every model, document it thoroughly enough to be recreated by a third party, validate and approve before deployment, then monitor and revalidate for drift. Generative frameworks extend SR 26-2 concepts into layered controls covering approved use, risk tiers, and evidence.

How it works

Model risk for AI in banking extends a discipline supervisors already know: the risk that a model produces wrong outputs or is misused. A central financial authority's December 2024 information paper on AI model risk management set out good practices from a thematic review, organized around governance and oversight, identification, inventory, risk materiality assessment, development and deployment standards, data management, validation, and deployment monitoring with change management, with dedicated chapters for generative AI and third-party AI. The how-it-works view matters because none of those stages is optional paperwork; each one produces an artifact that the next stage consumes, and the loop only holds if every link exists.

The loop starts with registration and inventory. A model that is not inventoried cannot be classified, tested, or monitored, so the first control is a complete register of models in use, including third-party and embedded ones. Risk materiality assessment then ranks them: a credit decisioning model and an internal drafting assistant carry very different exposure, and the resulting tier determines how much validation, documentation, and monitoring each receives. Operating guidance from a security body makes the sequencing explicit, since inventory scopes classification, classification decides testing, testing baselines monitoring, and monitoring feeds evidence, which is why AI governance works as a loop rather than a checklist and why bolting on one control at a time keeps failing audits.

Development standards carry the documentation burden. Regulators evaluate models through documentation that must be detailed enough for a knowledgeable third party to recreate the model without access to the development code, and the same record improves communication with auditors and bank management. Structured artifacts make that achievable: model cards capture purpose, training data, performance, limitations, explainability, and adversarial robustness; data sheets document datasets, assumptions, preprocessing, architecture, and validation logic; and risk cards act as registers per model with risk categories, severity, likelihood, impact, and mitigations. That linked structure is the core of a published model risk management framework for modern machine learning and generative systems.

Validation is the independent challenge before anything ships. It tests the model against expected behavior, evaluates it for bias and error-prone integration, and applies a formal approval process with criteria covering business need, architecture fit, and regulatory expectations before deployment. Implementation risk sits here too, because a soundly developed credit scoring model can still be wired into the loan processing system incorrectly and produce inaccurate assessments or unfair denials, so approval covers the deployment path and not only the mathematics. Revalidation schedules detect drift over time, and output monitoring watches for biased, toxic, or non-compliant content, particularly in customer-facing scenarios.

For generative AI the frameworks extend rather than replace the classics. A recent research paper proposes a generative AI control framework compatible with SR 26-2, translating model risk management concepts into a layered control structure for generative usage: approved-use boundary determination, risk-tier assignment, and evidence generation around workflows that both analyze and communicate. The surrounding activities matter as much as the score, since an institution must document model assumptions, monitor performance, explain results, interpret policy requirements, prepare escalation rationales, and support customer and regulator communications, and all of it becomes part of the model's risk record.

Third-party AI receives its own treatment because the bank's exposure can change without any release on its side. Institutions maintain an inventory of vendor models with use case, risk rating, and compliance exposure; require disclosures on model architecture, training sources, and system limitations; and put AI-specific terms in contracts covering data use, audit rights, incident response, and model updates. Change detection closes the loop: when a vendor swaps the underlying model, the classification, testing, and monitoring baseline must be reopened, and periodic vendor reviews re-examine the models that influence credit decisions.

Run end to end, the machinery lets a bank adopt agents without betting the balance sheet on them. The Bank of England's model risk management principles for banks, the Financial Stability Board's sound practices for responsible AI adoption, and supervisory guidance across jurisdictions converge on the same shape: clear accountability, complete inventory, risk-tiered validation, continuous monitoring, and documentation that satisfies a knowledgeable reviewer. Applied to agentic systems, the same loop governs the models that power them, so the bank's exposure to a wrong or misused model is measured, bounded, and evidenced rather than discovered in an incident.

Frequently asked questions

What does AI model risk management actually cover?

The full life cycle of models that can be wrong or misused: governance and oversight, identification and inventory, risk materiality assessment, development standards and data management, independent validation, and deployment monitoring with change management, including generative and third-party models. The practical output is a set of artifacts, an inventory entry, a risk tier, validation evidence, and monitoring baselines, each consumed by the next stage so the program behaves as a loop rather than a stack of documents.

Why is documentation treated as a control rather than a chore?

Because supervisors evaluate models through it: documentation must be complete enough that a knowledgeable third party could recreate the model without access to the development code. Structured artifacts make that achievable, with model cards covering purpose, data, performance, limitations, and robustness, data sheets covering datasets and validation logic, and risk cards acting as per-model registers of categories, severity, likelihood, impact, and mitigation. The same record is what auditors and bank management read.

How do the classic banking frameworks apply to generative AI?

They extend rather than get replaced. Recent research proposes a generative AI control framework compatible with SR 26-2 that turns model risk management concepts into layered controls for generative usage, including approved-use boundaries, risk-tier assignment, and evidence generation. The institution's obligations reach beyond the score to documenting assumptions, explaining results, preparing escalations, and supporting communications, so the model record grows to cover everything around the model as well as the model itself.

What is different about vendor and third-party AI models?

The exposure can change without any release on the bank's side, since a vendor may swap the underlying model quietly. The controls are an inventory of third-party tools with use case, risk rating, and compliance exposure, required disclosures on architecture and training sources, contract terms covering data use, audit rights, incident response, and updates, and change detection that reopens classification, testing, and monitoring whenever the underlying model moves. Periodic reviews re-examine the models that influence credit decisions.

Sources

  1. mas.gov.sg
  2. ey.com
  3. fsb.org
  4. bankofengland.co.uk
  5. abstracta.us
  6. kaufmanrossin.com
  7. aigl.blog
  8. cloudsecurityalliance.org
  9. aws.amazon.com
  10. netbankaudit.com
  11. arxiv.org
  12. cranium.ai
  13. metricstream.com
  14. - federalreserve.gov
  15. - federalreserve.gov
  16. - ithandbook.ffiec.gov
  17. - artificialintelligenceact.eu
  18. - eur-lex.europa.eu