SR 26-2 Model Scope for LLM Feature Extraction: Testing the LLM Step and the Downstream Model Separately

What changed: the April 2026 guidance and its generative AI carve-out

The OCC, Federal Reserve Board and FDIC issued updated interagency model risk management guidance on April 17, 2026. It is published by the Federal Reserve as SR 26-2. It supersedes SR 11-7 and SR 21-8, and the FDIC rescinded FIL-22-2017 and FIL-27-2021. Any internal procedure still built on those issuances needs to be rechecked.

Footnote 3 of the guidance reads: "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance." On May 1, 2026, Governor Bowman said the guidance now applies narrowly to traditional models and basic AI applications. The agencies also said on April 17 that they plan to issue a request for information on model risk management and banks' use of AI, including generative and agentic AI. Scoping decisions made today should be revisited when that request, or any further guidance, appears.

This changes the question the brief's premise assumes. The issue is no longer only whether an LLM component meets the model definition. It is also whether the component is a generative AI model that the guidance excludes, and what the bank does about controls and evidence when it is.

Apply the definition to each component

SR 26-2 defines a model as a complex quantitative method, system or approach that applies statistical, economic or financial theories to process input data into quantitative estimates. The definition excludes simple arithmetic calculations, such as those in spreadsheets, and deterministic rule-based processes and software where no statistical, economic or financial theories underpin their design or use. Each element can be applied separately to the LLM step and to the downstream model.

For the LLM step, the first test is footnote 3. The agencies' wording ties scope to what a component is: a generative or agentic AI model, a traditional model, or a basic AI application. It does not turn on whether the component sits upstream or downstream. Whether an LLM used only for extraction counts as a generative AI model, and how the exclusion treats such a component inside a pipeline that ends in a conventional model, are interpretive judgments. The bank should make them explicitly, record the reasoning and name who approved it.

For the downstream model, ignore the LLM at first and ask the ordinary questions. Is the method complex and quantitative? Do statistical, economic or financial theories underpin it? Does it produce quantitative estimates? A conventional scorecard or statistical model that meets those elements remains a model whatever feeds it. Whether structured features pulled from text are themselves quantitative estimates is a further judgment to document, not one to assume.

Check any post-processing as well. A deterministic rule that cleans or maps LLM output is excluded from the definition only where no such theories underpin its design or use.

Reliance and materiality shape the risk judgment

The guidance says model risk depends on the nature, scale and use of models in relation to the associated business risks. It also says the guidance is most useful for models supporting significant business lines, operations, services and functions, and that practices suited to one institution may not suit another with a different risk profile or model purpose.

Two facts therefore belong in the scoping record. The first is how much of the downstream model's input comes from LLM output, wholly or partly, and whether the model could run without it. The second is how material the affected decisions are. An extraction step that fills a few optional fields for a low-stakes workflow is a different case from one that supplies most of the inputs to a credit or compliance decision.

Institution size also matters. The FDIC states that the guidance generally does not apply to models used by banking organizations with total assets of $30 billion or less, to the extent they lack significant model risk exposure. Confirm where your institution stands and which supervisor's reading applies to it before relying on any exclusion.

What the determination means for model risk management evidence

The guidance covers model risk and its management, model development and use, validation and monitoring, governance and controls, and vendor and third-party products. If the downstream model is in scope, those areas apply to it, including the third-party aspects where a vendor supplies it. Whether they extend to an excluded generative AI component is not something the exclusion settles. Treat that as an institutional risk judgment.

The guidance also says supervisory action may result from violations of law or unsafe or unsound practices stemming from insufficient management of model risk. Because that statement is general, an exclusion from scope is not a reason to hold no evidence. A sound position is to keep internal controls for the LLM component and be able to show what it was permitted to do and what it did in production. This is a governance judgment, not a sourced examiner requirement.

In practice, a bank can expect two evidence sets. One is the scoping record: the memo, the approvals and the rationale for each component. The other is operational evidence: the permitted behavior of the LLM step, and records of its actual interactions. If the downstream model is in scope, validation and monitoring evidence for that model sits alongside both. Operational records inform that work. They do not replace it.

Where runtime policy enforcement fits

Trussed is a runtime control product. It enforces policy in the execution path on individual LLM interactions and tool calls: allow, redact, block, or escalate for human approval. Enforcement happens at runtime, not through review after the fact, and it produces audit-ready evidence that is mapped to frameworks including NIST, NAIC, NYDFS, HIPAA and the EU AI Act.

For an LLM extraction step, this can support the documentation of a scoping determination in a bounded way. A policy can state what the component may receive, such as redacting certain fields before they reach the model. The resulting records show how each interaction was handled. That helps answer the examiner question of what the component was allowed to do and what it did.

The limits are specific. Trussed does not provide model lineage, dataset provenance or end-to-end tracing of workflows, and its evidence is not validation of the LLM or the downstream model. Financial services is not a proven vertical for Trussed. Insurance is the regulated financial sector where it is deepest, and its production experience spans insurance, pharma, healthcare and education. Treat any use here as supporting evidence for your scoping file, not as a scoping decision.