Applicability of SR 26-2 to Bank Holding Companies and Nonbank Subsidiaries

Start with the letter's stated applicability

Scoping SR 26-2 begins with the text of the letter, not with assumptions carried over from earlier guidance or from other regulatory regimes. The text should answer four questions. Who are the addressees? Does the letter contain asset-size thresholds or other tailoring? How does it treat nonbank subsidiaries of bank holding companies? How does it relate to the model risk management guidance an institution already follows?

Where the letter answers a question directly, the scope decision is a matter of reading it accurately. Where it is general or silent, the decision becomes an interpretation, and that interpretation belongs with legal and supervisory-relations teams, not with the team building controls. Supervisory guidance is also distinct from regulation, so its legal weight should be confirmed by counsel. Examiners still assess risk-management practices regardless of that characterization.

One discipline matters throughout: do not import thresholds from unrelated rules. Other regimes use asset criteria for their own purposes, and none of them define who falls within SR 26-2. This guide is not legal advice.

The consolidated view of a holding company

The Federal Reserve's Bank Holding Company Supervision Manual states that the Federal Reserve has supervisory and regulatory authority for all bank holding companies. It also sets out objectives for supervising nonbank subsidiaries: identify significant nonbank activities, evaluate the risk-management practices of the parent and significant nonbank subsidiaries, and evaluate the degree to which nonbank risks may threaten subsidiary depository institutions, including through the transmission of legal or compliance risk.

The supervisory activities listed for significant nonbank subsidiaries begin with understanding the holding company's business lines and legal entity structure, and identifying nonbank subsidiaries whose activities present material financial, legal or compliance risk to the consolidated entity. For AI scoping, this means examiners are likely to think in terms of legal entities and how risk moves between them. A model or AI system in an affiliate can matter to the consolidated organization even if the affiliate's own scope status is unresolved.

This is supervisory context, not a statement of what SR 26-2 requires. It does explain why an inventory organized by legal entity is the right starting artifact.

Entity-by-entity scoping questions

The table frames what to resolve and what to hold for each entity type. The resolution of each question comes from the letter and from legal interpretation.

• Holding company: Resolve: whether it is an addressee and what consolidated accountability looks like. Hold: central policy definitions, an inventory of AI use cases across the group, and the documented scoping decisions for every entity.
• Insured depository subsidiaries: Resolve: the basis of coverage, any size tailoring, and which model risk guidance they follow today. Hold: model and AI inventory, validation records and interaction-level evidence for systems in production.
• Nonbank subsidiaries: Resolve: whether and how the letter reaches them, directly or through the parent. Hold: the entity's activities, the AI systems it runs, and a rationale for any exclusion, including the basis for it.
• Entities with material risk to the group: Resolve: whether their activities present material financial, legal or compliance risk to the consolidated entity. Hold: priority status in control mapping, whatever the final scope answer.

What adjacent scoping patterns do and do not tell you

Other Federal Reserve and interagency materials show how scoping questions tend to be framed. They are analogies and do not determine SR 26-2's scope.

Appendix F to Part 225, the information security guidelines, applies to customer information maintained by or on behalf of bank holding companies and their nonbank subsidiaries or affiliates for which the Board has supervisory authority, with listed exceptions for brokers, dealers, persons providing insurance, investment companies and investment advisors. It defines subsidiary to exclude insured depository institutions and their subsidiaries. The lesson is that entity categories can be carved in or out, and the carve-outs differ by regime, so each entity should be classified individually rather than by assumption. The same appendix lets a holding company include a subsidiary in its own program or have the subsidiary run a separate one, requiring coordination but not uniform policies. In the 2001 preamble, the Board said the holding company is ultimately responsible for its subsidiaries' compliance.

The FDIC's 2017 adoption of model risk guidance previously issued by the FRB and OCC shows size-based tailoring: it was not expected to pertain to FDIC-supervised institutions under $1 billion in total assets unless model use is significant, complex or poses elevated risk. That is prior guidance from a different agency. It suggests that tailoring, where it exists, may combine size with the nature of model use, but whether SR 26-2 does so must come from its text.

Where runtime enforcement supports examiner evidence

Model risk guidance of the kind the FDIC adopted in 2017 describes an effective framework as including disciplined, well-documented development and effective validation. Those processes address what a model is and how it was tested. They say less about what an AI system was permitted to do, and did, in a given interaction. That is the gap runtime enforcement addresses, and it complements validation and inventory rather than replacing them.

Trussed enforces policy in the execution path, across models, agents, tools, MCP servers, AI sites and coding assistants. Enforcement is applied to individual interactions and tool calls, and the outcomes are allow, redact, block, rate limit or escalate for human approval. Because policy is enforced at runtime, not reviewed after the fact, each decision produces audit-ready evidence as a by-product of enforcement. That gives a risk team a record of what the AI was allowed to do and what happened.

For entity-level scoping, the useful design properties are practical ones. Evidence should be filterable by legal entity, each record should carry the policy version that governed it, and retention and data residency should be configurable to the institution's needs. Trussed deploys self-managed in the customer cloud with configurable retention and data residency. Trussed does not provide model lineage or dataset provenance, so those records need to come from other parts of the model risk program.

Trussed has customers in insurance, pharma, healthcare and education, and none of that should be read as financial-services deployment experience. The relevance here is the evidence pattern: interaction-level records, organized by entity, that a risk team can produce when an examiner asks how a specific AI system was controlled.