Approved AI Model Catalog for Banks: Contents and Approval Authority Under the 2026 MRM Guidance
An approved AI model catalog for banks should record, for every model in use, its identity and version, vendor and hosting details, intended and prohibited use, risk tier, validation status, data handling constraints, owner, conditions of approval, and review dates. Approval authority should scale with risk tier, with model risk management, an AI risk committee, information security, third-party risk, privacy, compliance and the business owner each holding defined roles. The catalog is only a control if its status is enforced when AI is invoked.
Minimum catalog contents
- Identity and version: unique ID, model name, version, and the prompt, fine-tuning, configuration and provider settings the approval covers.
- Vendor and hosting: internal or third-party, embedded or standalone, hosting location, and the provider update and change conditions.
- Use boundaries: intended use, prohibited use, and the rule for re-approval when use extends beyond the original application.
- Risk and validation: risk tier, validation status, validator, and the findings and limitations that carry into approval.
- Data handling constraints: permitted data classes, redaction requirements, retention and residency limits.
- Ownership and conditions: accountable business owner, technical owner, conditions of approval, approval date, review date and current status (approved, conditional, blocked).
What changed: anchor the catalog to the 2026 guidance
Many bank AI policies still cite SR 11-7 and OCC Bulletin 2011-12. As of November 2026, the OCC, Federal Reserve Board and FDIC have issued revised interagency model risk management guidance to clarify principles, set out a risk-based approach, and rescind prior guidance. The OCC bulletin (2026-13) is dated April 17, 2026 and lists OCC Bulletin 2011-12 among the rescinded items. The Federal Reserve publishes the revised guidance as SR 26-2, with an attachment titled Supervisory Guidance on Model Risk Management. Policies written around the older documents should be re-anchored to the revised text.
Two features of the revised guidance shape catalog design. First, it expects controls commensurate with the magnitude of model risk and says appropriate practices differ by bank and by model purpose. Second, it says clear roles and responsibilities benefit model risk management, with governance that delineates who is responsible for development, validation and ongoing monitoring. Neither point prescribes a catalog field list or an approver mapping. The structure below is a design standard that fits those principles, not a regulatory requirement.
Scope needs an explicit decision. A Fed-hosted copy states that the principles apply to traditional statistical and quantitative models and to non-generative, non-agentic AI models. The OCC bulletin separately describes generative and agentic AI as novel and rapidly evolving. A separate OCC supervisory framework says certain model risk management principles may be instructive for such systems, with the scope, frequency and independence of controls adapted to consequences, non-determinism and reliance on third-party or continuously updated models. Decide which policy governs generative and agentic entries, and document that decision in the catalog standard.
Tier the approval path by risk and materiality
A single approval path either slows low-risk use or under-reviews high-risk use. Tiering by use-case consequence, customer and financial impact, and reliance on third-party or continuously updated models lets effort follow risk, consistent with the commensurate-controls principle.
A workable pattern has three paths. For high-tier models, such as those supporting significant business lines or customer-affecting decisions, validation completes before approval and a model or AI risk committee grants it. For mid-tier models, model risk management approves under delegated authority and reports additions to the committee. For low-tier models with predefined conditions, the business owner registers the use and model risk management confirms the tier. In every path, validation and effective challenge stay independent of development and use. The guidance describes effective challenge as critical analysis by objective experts, applied throughout the model lifecycle.
Define who may approve conditional approvals and exceptions, and how unapproved-model findings escalate. The board and senior management remain accountable. OCC testimony has said that using a third party does not diminish the responsibility of the board and management, so vendor and embedded models belong in the catalog and carry the same approval path as internal ones.
Approval authority: a RACI-style mapping
The mapping below is a proposed design. Adjust it to your organization, but keep one accountable approver per tier and make every sign-off recorded in the catalog entry.
- Model risk management: Assigns tier, oversees validation and effective challenge, approves mid-tier additions, confirms low-tier registrations.
- Model or AI risk committee: Accountable for high-tier approvals, conditional approvals and exceptions; receives reports on mid-tier additions and unapproved-use findings.
- Information security: Consulted on hosting, access, and data flow; signs off before any tier moves to approved.
- Third-party risk: Responsible for vendor due diligence, contract terms and ongoing oversight scaled to criticality, including embedded vendor models.
- Privacy and compliance: Consulted on data handling constraints, prohibited use and regulatory fit; can hold an entry at conditional.
- Business owner: Requests the addition, owns the intended use, and is responsible for operating within conditions and for periodic recertification.
Make the catalog enforceable at runtime
Examiners tend to find the gap between the approved list and the models actually in use. A catalog kept as a document cannot close that gap, because nothing stops a team from calling an unlisted model or using a listed one outside its conditions. The catalog should instead be a machine-readable input to a policy decision made when AI is invoked: an approved entry is allowed, a conditional entry is allowed with constraints such as redaction of specified data classes, and a blocked or unlisted model is stopped.
This is a design approach rather than a supervisory requirement, and it has a few practical consequences. Status changes must propagate quickly, so a revoked approval takes effect without waiting for the next review cycle. Version and configuration identifiers must be present in the entry so a change in prompt, provider or data can move the status to conditional pending re-review. Each runtime decision should record which entry was matched, what its status was, and what action was taken. That record is what lets a risk team compare models in use against the approved list, and it also surfaces unapproved use that must be reconciled and escalated.
Where Trussed fits
Trussed enforces policy in the execution path across models, agents, tools, MCP servers and coding assistants, with decisions to allow, redact, block, rate limit or escalate for human approval. Catalog status can serve as an input to those decisions, so policy is enforced at runtime, not reviewed after the fact. Each enforcement action produces audit-ready evidence, which gives a regulated enterprise a record of what its AI was allowed to do and what it did. Trussed does not provide model provenance, model lineage or dataset provenance.
Insurance is the regulated financial sector where Trussed is deepest. Its operating record includes more than 18 months of continuous production in a regulated enterprise, 99.95% uptime on live production traffic, and more than 100 models under one policy plane. It can be deployed self-managed in the customer cloud, with configurable retention and data residency. The catalog standard and approval structure described above stand on their own and do not depend on any particular enforcement product.
From catalog entry to examiner evidence
Catalog entry
Versioned record with tier, owner, conditions and review date
Approval path
Sign-offs scaled to risk tier and materiality
Runtime decision
Allow, redact or block based on catalog status
Reconciliation
Models in use compared with the approved list
Turn the catalog into a control
Discuss how catalog status can drive allow, redact and block decisions in the execution path, and what audit-ready evidence that produces.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.