ArticlesModel risk

SR 26-2 Excludes Generative AI. Your Model Risk Obligations Don’t.

SR 26-2 excludes generative and agentic AI from its scope while the UK's SS1/23 pulls the same pipeline in. Here is the evidence set that survives both regimes.

Jordan Browne-Moore October 2026 12 min read

Contents · 7 sections
  1. What the old guidance left open
  2. The route decides who asks
  3. One pipeline, two regimes
  4. Is the pipeline a model
  5. The evidence set
  6. Dates
  7. Where this stops

The OCC’s news release on the new model risk guidance gives the generative AI exclusion in two sentences, tacked onto a paragraph about asset thresholds.

“In addition, generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance.”

Footnote 3 of the guidance itself runs to four.

“Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance. Nonetheless, a banking organization’s risk management and governance practices should guide the determination of appropriate governance and controls for any tools, processes, or systems not covered in this document. However, the principles described in this guidance apply to traditional statistical and quantitative models and non-generative, non-agentic AI models.”

The release keeps the first two sentences and drops the rest. The fourth sentence is the one that draws the boundary, and it draws it by architecture, with use and materiality playing no part. A gradient boosted scorecard that decides credit is in. A large language model that reads the document the scorecard is fed is out. Five pieces of commentary on the guidance sit in the collection I made on 14 October 2026, from an industry foundation, two consultancies, a content marketing site and a vendor, and none of them quotes it.

The third sentence leaves the bank to decide, through its own risk management and governance practices, what governance and controls generative systems need. That is weaker than a requirement and more demanding in practice, because a requirement tells you what adequate looks like and a relocation does not. OCC Bulletin 2026-13 quotes the same two sentences twice on one page and follows both with the asset threshold sentence. The third and fourth sentences of the footnote appear on neither summary page.

Neither summary page carries them, and the full text sits one link away in the attachment. Read from either summary page, the exclusion looks like an exemption, and at least one vendor post has gone further and described the guidance as bringing generative systems into scope, which the second sentence contradicts outright.

For the record, and dated. I read the exclusion as a scope boundary in May 2026, in a piece that already cited SR 26-2, and two of the pages I checked again on 14 October 2026 were still keyed to SR 11-7.

What the old guidance left open

SR 11-7 ran from April 2011 and defined its subject without anticipating any of this. Under it, the open question about a language model in a credit pipeline was whether it counted as a model at all, and a validation function could spend a year on that without reaching a view. SR 26-2 supersedes and replaces SR 11-7 and SR 21-8, and footnote 3 answers the question for generative systems in the negative. The argument is over, and what replaced it is an obligation to specify. The bank now writes down what controls a system the agencies declined to specify for should have.

The route decides who asks

Take an LLM pipeline that feeds a credit decision, either by extracting fields from bank statements or by answering analysts’ questions in SQL. Whether you call a frontier API or host the model yourself decides who asks you for evidence about it, and in what form.

Call a frontier API and the pipeline is a purchased component. The bank’s procurement and supplier diligence process asks, and it asks for a vendor pack and a set of assurances about a system you did not build and cannot inspect. Host the model yourself and the pipeline is an internal system. The model risk function asks, and it asks for a validation file.

The obligation survives either route. It changes hands.

Section VII of the guidance, on purchased models, gets less attention than footnote 3 and says more about what to produce. It opens by acknowledging that components can be proprietary, so that a bank may not receive the code, data or methodology it would have had from a model built in house. Then it states its position.

“Nevertheless, the principles of model risk management remain applicable.”

And then the operative passage.

“Sound practice includes developing an understanding of the vendor model, including its conceptual soundness, design, development data, and performance. Similarly, sound practice involves conducting ongoing monitoring and outcome analysis to assess whether vendor models are accurate, remain fit for purpose, and continue to be reliable.”

Every frontier API a bank or its lending partner calls is a purchased product whose code, training data and methodology the buyer will never see, which is the condition the section was drafted for. One correction to the obvious reading of that, which I have made myself. Footnote 3 takes generative models out of the whole guidance, section VII included, so section VII binds a frontier API only by analogy. Sentence three makes the analogy likely, because a bank deciding what controls are appropriate for an excluded system will reach for the nearest written ones, and section VII is nearest. A second route in does not depend on analogy at all. The consuming credit model is in scope, its validation covers the quality of its inputs, and a generative pipeline supplying those inputs is reached through the consuming model whatever footnote 3 says about the pipeline.

Section VII names four things to understand, which are conceptual soundness, design, development data and performance. For a closed frontier model, three of the four go only as far as the vendor chooses to publish. Performance is the one that remains, and the only way to obtain it is to measure it on your own data.

One pipeline, two regimes

United States, SR 26-2 attachmentUnited Kingdom, SS1/23 as amended
Scope testArchitecture. Footnote 3 removes generative and agentic AI. The principles keep statistical models and AI models that are neither.Function. Paragraph 2.4 catches any quantitative method processing input data into output, including qualitative input. No technique is named.
PopulationMost relevant to organisations “with over $30 billion in total assets”, with a tailoring sentence reaching smaller ones.Firms holding internal model approval for regulatory capital. Credit unions, insurers and reinsurers are excluded.
InstrumentGuidance that “does not set forth enforceable standards or prescriptive requirements”, with supervisory action reserved for violations of law or unsafe or unsound practices.Expectations that LIAF01/26 clarified are “not conditions for internal model approval”.
Who asks, and for whatPurchased route, supplier diligence under section VII. Internal route, the bank’s model risk function. Footnote 3 sentence three governs the generative component on both.The framework directly. Inventory, tiering by materiality and complexity, validation, monitoring, and an annual self assessment against the five principles.
Date it bites17 April 2026, on issue. No compliance date, because there is no compliance.17 May 2024 for the policy, 23 April 2026 for the amendments, and 12 months from the grant of a new permission.

The populations stop this becoming a neat symmetry. A fintech lender is outside SR 26-2 because it is not a banking organisation, which is a different question from the $30 billion tailoring sentence, and its sponsor bank may sit on either side of that threshold. On the UK leg, SS1/23’s expectations reach only firms with internal model approval, and most UK banking partners of fintech lenders do not hold one. For those partners the statement is not binding and the question falls back to the firm’s own governance. A group running one pipeline across both regimes should expect two conversations about the same system, on different schedules, and should not assume either regime settles the other.

Is the pipeline a model

Without a limit, this argument turns every upstream data step into a model. A file transfer that moves a statement from one bucket to another touches data on its way to a credit decision and determines nothing about it.

The UK definition does the work here, so run it both ways. Paragraph 2.4 defines a model as “a quantitative method that applies statistical, economic, financial, or mathematical theories, techniques, and assumptions to process input data into output”, and adds that “Input data can be quantitative and/or qualitative in nature or expert judgement-based and the output can be quantitative or qualitative”. Paragraph 1.3 extends the framework to all models including vendor models, so a purchased component is inside the framework and not beside it.

For extraction, the in scope reading is the strong one. Fields pulled out of a KYC pack, a loan file or a bank statement become the input vector of the credit assessment. Monthly income extracted from a statement is not adjacent to the decision. It is the quantity the decision is computed from, and an error in it produces a wrong decision with nothing else in the chain having failed. Qualitative input, quantitative output, which is the combination paragraph 2.4 names. The out of scope reading has to argue that extraction is transcription, and it fails on any field the model had to normalise, classify or infer.

The scope question has to be run per use case. Banks are deploying these systems across credit underwriting, fraud detection, AML compliance, early-warning systems, customer service and regulatory reporting, and the answer differs depending on whether the system produces a quantitative output that feeds a decision or drafts and summarises for a human who decides. For query-based pipelines that return figures the database already computed, a reviewer can reasonably call the language model a retrieval interface rather than a model. For credit memo drafting and financial data extraction supporting underwriting, the case for scope is stronger. Each deployment has to be assessed on what it actually produces, not on the label of the technology.

Which is why the evidence set gets produced wherever the line lands. The cost of producing it is bounded and known. The cost of discovering mid diligence that your position rests on a contested reading is neither.

The evidence set

An accuracy figure is produced to be high. A validation file is produced to be attacked. A reviewer tests whether your numbers survive someone paid to find the gap between what the system is for and what the evidence shows, which means every number needs a stated definition, a sample, a date and a known weakness, and the weaknesses read better stated by you than found by them.

This list is my judgement, built from the supervisory principles. No regulator publishes one. The first three come from measurement and the other six from governance, and the companion piece on evaluation design is what produces the first three for either pipeline type.

1. Gold set provenance and construction. Who annotated it, against what written instruction, with what measured disagreement between annotators, refreshed when. Rejected: an evaluation set assembled by the team that built the pipeline, from documents the pipeline had already handled, with no record of how disputes were settled.

2. Specified tests. For extraction, accuracy at field level, format adherence, behaviour on document types never seen, and the definition of a correct answer for each field including units, rounding, period convention and absent values. For a query pipeline, the equivalence rule that decides when two result sets count as the same answer, implemented in both a strict and a containment variant with both numbers reported, and the resolved reading of the question graded separately against a written policy. Rejected: the word “tested” standing where a test specification belongs.

3. Monitoring thresholds and the logic behind them. The number that triggers action, who chose it, what evidence supports that number, and what happens when it trips. Rejected: a monitoring section that reports metrics and names no threshold.

4. The institution’s written specification of intended purpose. Footnote 3, sentence three, hands the institution the determination of appropriate governance and controls. A document stating what the pipeline is for, what it is not for, and which controls were judged appropriate and why is the direct answer to that sentence. Rejected: nothing on the record showing the determination was made at all.

5. Model inventory and tiering. The pipeline appears in the inventory with a materiality and complexity rating, the reasoning behind it, and its dependencies in both directions recorded. SS1/23 also expects responsibility for the framework to be allocated to a named senior manager, with the statement of responsibilities updated to match. Rejected: an inventory omission, which tells a reviewer the scope decision was never taken.

6. Data lineage and input governance. Where the documents come from, what the reference data is, how prompt and model versions are recorded against each output, and how an output traces back to the region of the source it came from. Rejected: aggregate accuracy offered by a team that cannot reconstruct one extraction end to end.

7. Independent validation. Somebody who did not build it, with the standing to say no, having reviewed it and written down what they found. Internal second line or external, and if external, with the engagement scope stated. Rejected: a review by the building team’s own manager.

8. Compensating controls where evidence is unobtainable. For a closed frontier component, three of the four things section VII asks about are not available. Name the gap and show what compensates, usually tighter output checking, narrower permitted use, and human review where a wrong value is expensive. Rejected: a vendor marketing pack offered as conceptual soundness evidence.

9. Named owners. A person for the pipeline, a person for the consuming model, and a person for the boundary between them. Failures cluster at the boundary nobody owns. Rejected: an ownership map with one name in all three places, or with the boundary row empty.

Which of the nine is the deliverable depends on what the measurement work finds. Where the gap turns out to be unwritten definitions, items 2 and 4 are what you ship. Where it survives the definitions being written down, the harness and the evidence file are what you ship. The companion piece sets out the audit that decides which.

Dates

  • 17 April 2026. The interagency model risk guidance, attached to SR 26-2 and to OCC Bulletin 2026-13. Effective on issue.
  • 23 April 2026. LIAF01/26 and the amended SS1/23, which pulled stress test practices toward the SS1/23 framework and clarified that the expectations are not conditions for internal model approval.
  • 16 November 2026. The proposed interagency guidance on third party risk management, 91 FR 58536, published 15 September 2026 by the OCC, the Federal Reserve Board, the FDIC and the National Credit Union Administration, carries a DATES field reading “Comments must be received on or before November 16, 2026.”
  • Pending. The bulletin states that “the agencies plan to issue in the near future a request for information that addresses model risk management generally and considers, in particular, banks’ use of AI, including generative AI and agentic AI and AI-based models.” The closing words take the request past generative systems to every AI model a bank runs. It has not been published as of 18 October 2026.
  • 2 December 2027. Obligations for Annex III high risk systems under the EU AI Act apply, following the deferral in Regulation (EU) 2026/1744. Article 50 transparency obligations were untouched by that regulation and have applied since 2 August 2025.

Annex III point 5(b) makes a system intended to evaluate the creditworthiness of natural persons high risk, so this reaches consumer lending and not corporate. Where the extracted fields are what the evaluation is computed from, the provider must decide and document whether the pipeline is a component of that system or a system of its own. The EU dates above rest on the EUR-Lex consolidated text, which carries its own notice that it has no legal effect.

Where this stops

The scope argument for a query pipeline is contested, and the section above says which way a reviewer may rule. The evidence set is practitioner judgement mapped onto supervisory principles and no regulator has endorsed it. The commentary collection of 14 October 2026 was not exhaustive. One property of these systems sits outside all of the above and this piece does not measure it, which is that a pipeline querying a credit database holds whatever permissions it was given, and hosting the model yourself moves that exposure onto your own control plane.


Sources