BSA/AML Model Validation After SR 26-2?

SR 26-2 reframed the guidance, but it didn’t change the regulation, and it didn’t change the best practices.

At ARC, we have been validating BSA/AML models for over 17 years, since the company was founded — through SR 11-7 and OCC Bulletin 2011-12 in 2011, through SR 21-8 in 2021, and now SR 26-2. Independent testing of BSA/AML systems was best practice then and continues to be. SR 26-2 changes the nuances of how testing is scoped and evidenced, and that effect is most visible at institutions with mature model risk programs. For everyone else, the broad strokes are still the same. What constitutes sound BSA/AML testing has not changed.

What Changed and What Didn’t

On April 17, 2026, the Federal Reserve issued SR 26-2, promulgated in parallel as OCC Bulletin 2026-13 and FDIC FIL-15-2026. It supersedes SR 11-7 (2011) and SR 21-8 (2021), sets an applicability threshold of $30 billion in total assets, narrows the definition of “model,” and carves out generative and agentic AI. It preserves conceptual soundness, outcomes analysis, and ongoing monitoring, and states that “model purpose, together with model exposure, determines model materiality.”

SR 11-7 didn’t invent model governance, and SR 21-8 didn’t create the duty to validate BSA/AML systems. SR 21-8 even said so on its face: it “does not alter existing BSA/AML legal or regulatory requirements, nor does it establish new supervisory expectations.” The regulatory obligation and best practices pre-existed the guidance, and SR 26-2 didn’t rescind them.  

What SR 26-2 left untouched: the BSA program requirement at 31 U.S.C. § 5318(h) and 31 C.F.R. § 1020.210, and the independent testing pillar codified at 12 C.F.R. § 208.63(c)(2) (Federal Reserve), § 326.8(c)(2) (FDIC), § 21.21(d)(2) (OCC), and § 748.2(c)(2) (NCUA). SR 26-2 notes in its own first footnote that while it sets no enforceable standards, “supervisory action may result for any violations of law or unsafe or unsound practices stemming from insufficient management of model risk.”

Does 26-2 Provide Relief?

Our clients’ reactions to the guidance reflect the ambiguity of this moment. Most find themselves waiting and watching to determine what to continue, what to reconsider and where they can economize.  

The real answer is that it depends. For those with well-developed model risk management, SR 26-2 means revisiting risk-based tiering, categorization and scheduling. But many institutions, still in the process of developing a solid risk assessment and governance program, should approach this differently. The OCC indicated: “These actions build upon the OCC’s ongoing efforts to tailor its supervisory framework to reduce unnecessary burden and promote risk-based examination across institutions of all sizes.” Vice Chair for Supervision Michelle Bowman wrote: “Over time, supervisors expanded the scope of the previous guidance beyond its original purpose to apply it in unintended ways.”  This was specifically about AI but relates to risk management generally. She’s talking about an intelligent best practice regime turning into a routine checklist. Validation was always essential to sound risk management.

Why We Still Need to Validate BSA Systems

Many of the largest recent BSA failures point to testing deficiencies. At U.S. Bank, assessed $613M in 2018, independent testing would have surfaced alert volume caps, the absence of coverage for non-customer Western Union activity, and below-the-line samples in which roughly half the transactions proved suspicious. At Standard Chartered’s New York branch, assessed $300M in 2014, testing would have revealed that the surveillance system was failing to detect a large number of potentially high-risk transactions. And at TD Bank, assessed $3B in 2024, a systematic review of exclusion decisions could have revealed that domestic ACH, most check activity, and other transaction types were excluded from automated monitoring, leaving $18.3 trillion unmonitored from 2018 to 2024.

BSA/AML and OFAC monitoring and screening systems require testing whether they are categorized as models or not. Under SR 26-2, some systems’ thresholds, watchlists, and fuzzy-match tolerances may be considered “deterministic rule-based processes.” If that is the conclusion, they may fall outside model risk management but remain fully within the scope of the BSA independent testing obligation. Worth noting: the same analysis can put a system in scope rather than out. Machine learning that tunes thresholds, segmentation, or match tolerances is itself a model, and SR 26-2’s carve-out reaches only generative and agentic AI.

Regardless of classification, we would expect to see above-the-line and below-the-line threshold testing; data completeness, accuracy, and lineage validation; scenario coverage assessed against the institution’s risk assessment; alert-to-case-to-SAR analysis; and watchlist and fuzzy-matching tuning.

Why We Need to Test AI in Particular

SR 26-2 carves out generative and agentic AI, but the carve-out is about timing.   The technology is moving faster than the guidance can be written, and the agencies said so in footnote 3: these models “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.”  

Banks are already using generative and agentic tools for alert triage, SAR narrative drafting, adverse media screening, and CDD summarization. The limitations those functions inherit are well documented outside banking, and three of them bear directly on testing:

Hallucination. A public database maintained by HEC Paris researcher Damien Charlotin catalogs court decisions in which a party relied on AI-hallucinated material and a court responded — over 2,000 decisions worldwide, up from roughly 200 in mid-2025. These are only the filings a judge caught. The systems produced confident, well-formed citations that did not exist. Historical testing of outputs as well as analysis of structural safeguards are both essential when relying on AI processes.

Drift. Stanford and UC Berkeley researchers evaluated the March 2023 and June 2023 versions of GPT-4 on identical tasks and found accuracy on prime-versus-composite identification fell from 84% to 51% over three months, with no action by the user and no notice from the vendor. The behavior of the same service can change substantially in a short period. It’s worth noting the applicability of on-going monitoring as a strategy for addressing drift.

Unbounded Action. In July 2025, an AI coding agent on Replit’s platform executed destructive commands against a live production database during an explicit code freeze, deleting records for roughly 1,200 executives and 1,200 companies. It then generated thousands of fabricated user records and told its operators that rollback was impossible. The false account of what had happened nearly made the loss permanent. Authorization boundaries and independent audit logging are the controls at issue here.

AI tools performing important functions within a BSA program are subject to the independent testing requirements that apply to that program. They are temporarily carved out of testing guidance under SR 26-2, but not from 12 C.F.R. § 208.63(c)(2).

Impact by Institution

Above $30B, we would focus first on the materiality framework and on how systems are tiered within it. In our experience, BSA/AML systems may warrant higher tiering under SR 26-2 than they often received before. Below $30B, the activity-based carve-in is the thing to examine. The guidance may still reach institutions with “activities outside the scope of traditional community banking,” with BaaS and fintech partnerships the clearest case. Credit unions were never covered. NCUA was not a party to SR 26-2, and § 748.2(c)(2) is unaffected. For all, the independent testing obligation and the FFIEC exam expectations apply regardless.

SR 26-2 asks institutions to make their own judgments about materiality, scope, and testing cadence, and to defend those judgments to examiners. That is a harder standard than following a checklist.