A Regulator Says Use AI. Another Says the Framework Is Not Ready.

In April 2026, the federal banking agencies retired SR 11-7, the model risk guidance that had shaped how banks validate AML models for 15 years. Its replacement, OCC Bulletin 2026-13 and SR 26-2, places generative and agentic AI outside its scope, calling them novel and rapidly evolving. Existing risk practices should still guide the controls banks apply, and a request for information on AI is coming.

The same month, FinCEN proposed a fundamental reform of AML/CFT programs, shifting the standard to demonstrated effectiveness and promising no added enforcement risk for responsible experimentation with machine learning and generative AI.

One regulator is encouraging banks to bring AI into their financial crime programs. The prudential supervisors have said the framework for that AI has yet to be written. 

I have seen a version of this tension in traditional AML model governance, especially around transaction monitoring tuning. When audit or model risk management reviews a tuning change, they ask for much more than the final threshold: the data and population used for testing, the methodology and assumptions, before-and-after alert volumes, coverage and false-positive considerations, the SME judgment applied and who challenged and approved the change. 

I have also seen parts of that trail rebuilt after the fact from analysis files, emails, meeting notes and earlier test outputs, because the rationale never made it into the final documentation.

The questions examiners will ask

The questions are foreseeable and are the ones Bank Secrecy Act (BSA) officers already field about models, stretched to fit systems that act.

Explainability: When an agent closes an alert after reviewing six sources, the bank has to show what it consulted, what it weighed and why, in a form an examiner can follow.

Accountability: An agent built from one vendor’s language model, another’s entity resolution and internal data connectors has diffuse hands on it. When the output is wrong, the bank owns the outcome, and someone inside it has to sign their name.

Human oversight. KPMG’s 2026 research with US banking leaders found human-in-the-loop design was the central question as teams tried to expand capacity without breaching AML and KYC oversight expectations. Human review can be a control or a rubber stamp, and examiners will want to know which is actually happening in practice.

And the audit trail, which underpins all three. If reasoning and actions are not logged at each step, none of the above can be reconstructed after the fact.

Ongoing effectiveness belongs on the list too: how the institution knows the process still performs as intended, what triggers recalibration and how changes are tested and approved, along with whether the agent’s inputs are complete, accurate and traceable. The emphasis shifts by function, toward scenario coverage, thresholds and escalation in transaction monitoring and toward list coverage, matching logic and false negatives in sanctions screening. In either case, a central question will be where autonomous decision-making ends and human review begins.

Financial Services

25 Years After Sarbanes-Oxley, How Is Fraud This Simple Still Possible?

by Anne Eberhardt

October 12, 2026

Lessons for lenders from two billion-dollar failures seemingly built on misstated assets, double-pledged collateral and inflated invoices

Read moreDetails

Controls to put in place now

None of these has to wait for guidance; each is an adaptation of something AML model governance already requires.

• Enforce the permission line in system access: An agent that drafts suspicious activity report (SAR) narratives should have no ability to file them, and one that recommends dispositions should not close alerts on its own. Transaction monitoring tuning already works this way. The data analytics team I oversee may test a threshold change or analyze a new scenario and build a recommendation from the results. Compliance management and model owners weigh those results against SME judgment before approving anything, and the approved change goes to production through change management, implemented by authorized IT staff or the vendor. I hold compliance systems administrator access and that still does not let me move a tuning recommendation into production myself. An agent that helps with the analysis should not pick up approval or implementation rights along the way. 
• Log reasoning at every step: A usable record captures what data the agent retrieved and from where, which tools it invoked and in what order, the intermediate conclusions it drew, the recommendation it produced and which version of the model, prompts and connectors was running at the time. The log has to be retrievable by alert or case during an exam or a lookback, which is a different standard from being technically stored somewhere. 
• Measure disagreement with blind challenger review: Transaction monitoring QA can supply the template. A reviewer independently works a sample of dispositioned alerts, looking at the transactions, customer activity and scenario logic and records whether they agree with the investigator. Disagreements that change the outcome count for more than differences in documentation, and recurring ones get traced to the investigator, the procedure or the scenario’s tuning. For an agent, sample its outputs on a set cadence, route them to investigators who cannot see its conclusion and track how often they disagree and in which direction. An agent that closes or de-escalates alerts a human would have escalated is the failure mode regulators care about. 
• Stage autonomy and make each expansion earned: When a new scenario or material enhancement comes into transaction monitoring, we validate it in a non-production environment against historical transaction data, often several months of it, enough to capture realistic volume and behavior. We check whether the scenario catches the activity it was designed for, review alert volumes and population coverage, analyze false positives and false negatives where they can be evaluated and compare results with the risk the scenario was meant to address. The analytics team, compliance SMEs, model owners and compliance management review the results, parameters get refined, and the change reaches production only after governance approval, through change management. An agent should clear the same gates. Define expected outcomes before testing begins, document exceptions and challenge and keep investigators reviewing its outputs after go-live before it is allowed a final disposition. 

Conclusion

Regulator-ready governance rests on documentation a non-technical examiner can follow and on change control that captures the reason a model changed. Those principles hold. Some things do not. Static validation on a frozen model assumes it behaves the same way on the same inputs, which a language-model-driven agent may not. Back-testing against historical alerts assumes the job is classification, when an agent’s job may include gathering the evidence a classifier would then use. And the model inventory, which assumes discrete assets, will strain under a system whose foundation model, tools, prompts and connectors each change independently. 

Federal Reserve Vice Chair for Supervision Michelle Bowman has called for an assessment of whether current AI guidance is fit for the future, acknowledging the revised model risk framework does not reach generative or agentic AI. When that guidance arrives, it will draw on what supervisors have seen work at institutions that took governance seriously during the gap.

Governance maturity is the throttle on how much autonomy a bank can safely hand to a machine. Teams that build it before deployment will be better positioned to scale agentic capabilities while less-prepared institutions may find themselves addressing governance and remediation gaps.