The Conventional Wisdom
Many believe you should wait for regulatory guidance before deploying AI in your AML program. The prudential supervisors retired SR 11-7 in April 2026 and excluded generative and agentic AI from OCC Bulletin 2026-13 and SR 26-2, labeling these systems as "novel and rapidly evolving." Without a published framework from examiners, it seems like you're navigating without a map.
This caution seems sensible. No one wants to justify an autonomous alert-closing system to the OCC without explicit regulatory approval.
Why You Should Act Now
You already have the framework. The controls for an agentic AI system aren't new; they're adaptations of the model governance you currently use for transaction monitoring and sanctions screening.
Consider when your team proposes a new transaction monitoring threshold. You don't just approve it. You review the testing methodology, sample population, alert volumes, false-positive analysis, SME judgment, and who challenged the recommendation. Compliance management then approves it, and authorized IT staff implement it through change management.
An AI agent recommending alert dispositions requires the same rigor, applied to different mechanics. Examiners will ask questions BSA officers already answer: What did the system consult? How did it weigh the evidence? Who approved the logic? What happens when it's wrong?
FinCEN is shifting the standard to demonstrated effectiveness and has promised no added enforcement risk for responsible experimentation with machine learning and generative AI. This is an invitation to act, not to wait.
The Evidence
Three control requirements in your AML governance program map directly to agentic systems:
Explainability translates to audit trails. When an agent closes an alert after reviewing data sources, you need a log showing what it consulted, what it weighed, and why. This is the same standard applied when an investigator documents their rationale. The difference is the agent's reasoning must be captured automatically at each step, retrievable by alert or case during an exam.
Accountability translates to permission boundaries. Your analysts can't push their own tuning changes to production. They test, recommend, and document. Compliance management approves, and IT implements through change management. An agent drafting SAR narratives shouldn't have filing rights. One recommending dispositions shouldn't close alerts autonomously. The permission line for human users applies to machine users.
Human oversight translates to blind challenger review. Your QA already does this: a reviewer independently assesses a sample of dispositioned alerts and records agreement with the investigator's conclusion. Disagreements that change the outcome are crucial. For an agent, route its outputs to investigators who can't see its recommendation, track disagreement rates, and trace recurring failures to the model, prompts, or data connectors.
KPMG's 2026 research with US banking leaders found human-in-the-loop design was central as teams expanded capacity without breaching AML and KYC oversight expectations. Your existing controls provide the answer.
What to Do Instead
Build governance before autonomy. Four practices will prepare your program for examiner scrutiny:
Stage the expansion. Validate new transaction monitoring scenarios against historical data in a non-production environment. Check if it catches target activity, review alert volumes, analyze false positives, and compare results with the risk profile. Apply the same gates to an agent. Define expected outcomes before testing, document exceptions, and keep investigators reviewing outputs after go-live before allowing final dispositions.
Log every decision point. Capture what data the agent retrieved, which tools it used, the intermediate conclusions, and which model and prompts were running. The log must be retrievable by alert during an exam.
Measure disagreement systematically. Sample the agent's outputs regularly and route them to investigators who work them blind. Track disagreement frequency and direction. An agent closing or de-escalating alerts a human would escalate is a failure mode regulators care about.
Separate recommendation from execution. Your data analytics team tests a threshold change and builds a recommendation. They don't implement it. Compliance management weighs the results against SME judgment, approves the change, and authorized staff push it to production. An agent assisting with analysis shouldn't have approval or implementation rights.
These aren't theoretical controls. They're the ones your AML governance program already enforces, applied to systems that act rather than classify.
When Conventional Wisdom Is Right
Some problems lack precedent. Static validation assumes a model behaves consistently with the same inputs. A language-model-driven agent may not. Back-testing against historical alerts assumes the job is classification, while an agent's job may include gathering evidence a classifier would use. Your model inventory, built for discrete assets, will strain under a system where the foundation model, tools, prompts, and connectors 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 doesn't reach generative or agentic AI. When that guidance arrives, it'll draw on what supervisors have seen work at institutions that took governance seriously during the gap.
If you're waiting for permission to address these problems, you're falling behind institutions that aren't. Governance maturity controls how much autonomy you can safely give a machine. Teams that build it before deployment will be better positioned to scale agentic capabilities while less-prepared institutions address governance gaps under exam pressure.
You don't need new rules. You need to apply the ones you already have.
