September 23, 2026

What "Agentic AI" Actually Means for Compliance Teams

Axle Team

Key takeaways

"Agentic AI" in compliance means software that works a case end to end: gathers evidence, reasons over it, documents a disposition, and hands defined decisions to a human. It is not a smarter rule or a chat window.
U.S. banking regulators drew a line in April 2026. The revised model risk management guidance (SR 26-2 / OCC Bulletin 2026-13) explicitly excludes generative and agentic AI and points institutions back to their own risk management and governance practices.
FinCEN's April 2026 program proposal would treat the use of "innovative tools such as artificial intelligence" as a factor weighing against enforcement, provided the tools demonstrate program effectiveness.
The governance question for agents is not "how autonomous is it" but "which decisions is it allowed to make, and how do we prove each one was made well."

Agentic AI compliance: what the term actually means

Agentic AI compliance means using AI systems that execute compliance work rather than just scoring it. An agent takes an alert or review, pulls the relevant customer, transaction, and third-party data, reasons through the evidence, writes a disposition with rationale, and escalates decisions it is not authorized to make. Humans set the decision rights, review samples, and own the outcome.

The definition matters because the phrase is used loosely. Vendors call anything with a language model "agentic." Examiners hear it and think of model risk management. Neither reading helps a BSA officer decide what to deploy or how to govern it.

The useful distinction is about work, not intelligence. A rule flags. A model scores. A copilot answers when asked. An agent completes a multi-step task, some steps involving judgment, and produces an artifact a reviewer can check.

Rules, models, copilots, agents: what changes at each step

Detection rule — What it does: Fires when a threshold or pattern is met. Who makes the decision: Analyst, after the fact. How it has been supervised: Tuning, threshold validation, scenario coverage.
Statistical or ML model — What it does: Scores or ranks risk. Who makes the decision: Analyst, using the score. How it has been supervised: Model risk management (SR 11-7 lineage, now SR 26-2).
Generative AI copilot — What it does: Summarizes or drafts on request. Who makes the decision: Analyst, who prompts and edits. How it has been supervised: Mostly general IT and third-party risk.
Agentic AI — What it does: Works the case: gathers, analyzes, documents, escalates. Who makes the decision: Agent within defined decision rights; human above them. How it has been supervised: Outside SR 26-2 scope; governed by the institution's own framework.

Where regulators landed in 2026

On April 17, 2026, the Federal Reserve, FDIC, and OCC replaced the 2011 model risk guidance. The new version defines "model" to exclude deterministic rule-based processes, and it carves generative and agentic AI out entirely:

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.
— Federal Reserve, FDIC, and OCC, Supervisory Guidance on Model Risk Management (SR 26-2 attachment), April 17, 2026

The same package rescinded the 2021 interagency statement on model risk management for BSA/AML systems. Fed Vice Chair for Supervision Michelle Bowman explained the reasoning in a May 1, 2026 speech:

Over time, supervisors expanded the scope of the previous guidance beyond its original purpose to apply it in unintended ways. We recognize that rapidly evolving and novel technologies like AI may require a different approach. The revised guidance now applies narrowly to traditional models and basic AI applications. Going forward, we expect other risk-management and governance practices to support adoption of generative and agentic AI in ways that will encourage ongoing innovation.
— Vice Chair for Supervision Michelle W. Bowman, Artificial Intelligence in the Financial System, May 1, 2026

Two things follow. An agent working AML alerts is not automatically a "model" needing SR 26-2 validation, though a statistical model inside the workflow still is. And the absence of a template is not the absence of expectations: the OCC's release the same day said the agencies plan a request for information that "considers, in particular, banks' use of AI, including generative AI and agentic AI and AI-based models."

FinCEN has gone further. Its April 7, 2026 proposal to rewrite the AML/CFT program rules lists, among the factors the Director would weigh before an enforcement or significant supervisory action, "whether the bank is employing innovative tools such as artificial intelligence that demonstrate the effectiveness of the bank's AML/CFT program" (FinCEN fact sheet). The operative words are "demonstrate the effectiveness." Deploying an agent earns nothing on its own.

What an agent actually does inside an AML workflow

An alert-handling agent runs a sequence any Level 1 analyst would recognize:

Intake. Receive the alert with the scenario or match that triggered it.
Evidence gathering. Pull the customer profile, KYC file, expected activity, prior alerts and SARs, and counterparty data.
Analysis. Compare observed to expected activity and known typologies; identify what explains the alert and what does not.
Disposition. Produce a written rationale and a recommended outcome: close with rationale, request information, or escalate for investigation.
Escalation and hand-off. Route anything outside its decision rights, such as a SAR filing decision, to a human with the full evidence file attached.
Record. Log every data source, inference, and output in a form QA and examiners can reconstruct.

Axle's agents follow this sequence for specific workflows: Rose for transaction monitoring alerts, Baley for screening alerts, and Sam for SAR narratives. The agent produces the case file; the institution decides what it may close and what it must escalate.

Governance that holds up when the MRM letter does not apply

Because SR 26-2 does not tell you how to govern an agent, your own framework has to. An agentic AI compliance program that examiners will recognize maps onto controls they already test:

Decision rights in writing. Which alert types, risk tiers, and outcomes the agent may finalize, and which require a human. This is the most important document.
Change control. Changes to what the agent can do go through change management with before-and-after testing, like a scenario threshold change.
Reconstructable case files. Every disposition shows the data used and the reasoning, so an examiner pulling a closed alert sees why without asking anyone.
Sampled quality review. A fixed percentage of agent-closed alerts reviewed by experienced staff, with error rates tracked and reported to the BSA officer.
Back-testing. Before go-live, run the agent on historical alerts with known dispositions and SAR outcomes. Measure agreement and study the disagreements.
Third-party risk. The vendor, the underlying language model, data handling, and access controls all fall under existing third-party risk management, which did not get narrower in 2026.
Effectiveness evidence. Track what FinCEN's proposal cares about: SAR quality, time to escalate genuinely suspicious activity, and how much analyst attention shifts toward higher-risk work.

One practitioner note. The most common first-year mistake is defining scope by "autonomy level" rather than by decision. A policy saying the agent operates at "level 2 autonomy" means nothing to an examiner. Write a decision matrix instead: this alert type, this risk tier, this outcome, agent may finalize; anything else, escalate. Then make the agent's case file identical in format to an analyst's, so QA can score dispositions blind. If reviewers can tell which files came from the agent, you have a format problem that will become an exam problem.

The AI readiness playbook walks through how to stage that decision matrix, and the ROI calculator helps size the analyst hours involved.

FAQ

Is agentic AI subject to model risk management requirements?

Under the April 2026 revised interagency guidance (SR 26-2 / OCC Bulletin 2026-13), generative and agentic AI are out of scope; institutions' own risk management and governance practices determine the controls. Traditional statistical models inside the workflow remain in scope, and an AI-focused request for information is planned.

Can an AI agent file a SAR?

An agent can gather evidence and draft the narrative, but the decision to file should sit with designated compliance staff. FinCEN's program proposal keeps the requirement for a designated AML/CFT officer responsible for day-to-day compliance, and most programs treat the SAR decision as human-only.

Do U.S. regulators allow agentic AI in AML compliance?

Nothing prohibits it. The 2018 interagency statement on innovation encouraged responsible pilots, and FinCEN's 2026 proposal would count effective use of AI tools as a favorable enforcement factor. The expectation is that the institution can show the tools work and are governed.

Next step

Start with the decision matrix, not the technology. Axle's platform overview shows how alert-handling agents, human decision rights, and case documentation fit together across monitoring, screening, and SAR workflows.

What "Agentic AI" Actually Means for Compliance Teams

More from the blog