AI in AML Programs: What Changes and What Doesn't

Last updated: October 2026

AI in AML programs changes how the work inside the program gets done: how alerts are triaged, how evidence is gathered, how narratives are drafted, how screening hits are reviewed and how monitoring is tuned. It does not change who answers for the program. The designated BSA officer, the institution's risk assessment, the decision to file a suspicious activity report (SAR) and independent testing all stay with people. Examiners will judge an AI-supported program on whether the institution understands its systems, can show where people decide, and can reconstruct every outcome.

TL;DR

• AI changes the work inside an AML program: alert triage, investigation, SAR narrative drafting, screening review, monitoring tuning and quality control.

• Accountability does not move. The regulation names an individual responsible for day-to-day compliance, not a system, and the SAR decision stays with the institution.

• When a person reviews every AI output, that review is quality control over the AI. Second-line quality assurance and independent testing are still needed on top.

• Model risk guidance changed on 17 April 2026. SR 26-2, OCC Bulletin 2026-13 and FDIC FIL-15-2026 replaced SR 11-7; machine-learning monitoring models are in scope, and generative and agentic AI are explicitly not.

• FinCEN withdrew its 2024 AML program proposal and proposed a new rule in April 2026. That rule is not final as of October 2026, so today's five-element rule still applies.

What does AI change in an AML program?

AI changes how the work inside an AML program is done, while accountability for that work stays where it was. Machine-learning models and AI agents can take on evidence gathering, first-pass review and drafting across alert triage, investigations, SAR narratives, screening review and monitoring. Every decision that carries regulatory accountability stays with a named person at the institution.

An AML program, under today's bank rule (31 CFR 1020.210), has five elements: a system of internal controls, independent testing, a designated individual or individuals responsible for coordinating and monitoring day-to-day compliance, training for appropriate personnel, and risk-based procedures for ongoing customer due diligence. The FFIEC BSA/AML Examination Manual calls the designated individual the BSA compliance officer. Most of what AI touches sits inside internal controls. The other elements describe who is accountable and how the program is checked.

The second column is where AI changes the work; the third is what an examiner will still expect a named person to own.

Some decisions stay human-owned in almost every program that uses AI today: final SAR determinations, high-risk onboarding approvals and sanctions true-hit confirmations. Speed is only part of the case for AI. A head of screening operations at a large bank said consistency of execution matters more, and that consistency is where high volumes make teams fall short.

The work that changes: triage, investigation, narratives and screening

Alert triage is where AI changes the work first. An AI agent can review an alert, gather the account and transaction context, and recommend a disposition with its reasoning, and the analyst then decides. Buyers tend to draw the line in the same place.

A BSA officer at a bank put it this way: "I don't think we'd ever get to the point where it does the case and the sar at least for the foreseeable future at this point. But I do know that a lot of folks are moving towards using agentic to replace the triage portion." How triage changes when an analyst signs off on every recommendation is the subject of our piece on L1 alert review.

In investigations, AI changes how evidence is assembled. Pulling transactions, prior alerts and customer records into one case record is work an AI system can prepare. What the activity means, and whether it is suspicious, stays the investigator's call.

SAR narrative drafting changes too, because a first draft can come from the case evidence. The roles around the narrative do not change. A BSA officer at a sponsor bank described the usual split: analysts investigate, prepare the SAR and write the narrative, and BSA managers do quality control, then sign it and submit it to FinCEN. AI can take on part of the drafting, while quality control, the signature and the filing stay where they were.

Buyers also ask whether automated drafting is legally sound, and QA teams set a firm limit. An AML quality assurance lead at a fintech said: "the last thing that I want to see is a whole bunch of AI generated sars with no human in the loop feedback within it." Our piece on SAR narratives covers how drafting and sign-off fit together.

Sanctions screening review changes in the same way. An AI system can lay out the evidence for and against a potential match, and the analyst confirms or clears it. A senior sanctions compliance officer at a bank pointed out that a system preparing the narrative for an analyst's decision is still part of the screening process, even if it does not screen names itself, so it still sits inside model risk. Our piece on screening alert review covers that workflow.

Not every program has a volume problem. A BSA officer at a community bank with no backlog of alerts, cases or SARs said that makes the case for AI harder to make. For a program like that, what AI offers is consistency and a better record of each decision.

Monitoring models and tuning: what changes in change control

Machine-learning monitoring models and AI-assisted tuning change how thresholds and scenarios move. They do not remove the approval step. In BSA, a rule change goes through review before it goes live: one BSA officer at a bank contrasted that with fraud teams, who change rules first and send them to model risk review afterwards.

The program owner still has to explain each change. A BSA program owner at a credit union described the test that comes once or twice a year: "I'm going to have to speak intelligently to be able to explain how my system is alerting and if it's changed in that time period, how I ensure that change is sound." Some teams would rather not tune alerts away at all. The BSA officer quoted earlier on triage wondered whether letting alerts fire and having an agent work them might go better with examiners.

Model risk guidance changed this year. On 17 April 2026 the Federal Reserve (SR 26-2), the OCC (Bulletin 2026-13) and the FDIC (FIL-15-2026) issued revised interagency model risk management guidance. It supersedes SR 11-7 and the 2021 interagency statement on model risk management for BSA/AML systems. A machine-learning transaction monitoring or screening model falls within the revised guidance's general definition of a model.

Generative and agentic AI do not. The OCC bulletin says: "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance." The agencies said they plan to issue a request for information on model risk management, with particular attention to banks' use of AI, including generative and agentic AI. As of October 2026 that request has not been issued.

Being outside the guidance does not leave an agent ungoverned. The same credit union program owner said independent model validation is guidance rather than law, but that it feels like a requirement examiners will raise if it is missing. How validation works after April 2026 is covered in our piece on AML model validation, and the choice between rules and models in our piece on monitoring rules and machine learning.

Quality control and quality assurance

When a person reviews every AI output before it moves on, that review is quality control over the AI. It is not the program's quality assurance. An MLRO at a payments company, thinking through how to present an agent deployment to a regulator, drew exactly that distinction: human review of each output is quality control over the agents, and a second-line quality assurance view is still needed on top.

Quality assurance over AI-assisted work should test in both directions. A director of AML compliance at a bank described QA over cleared cases that looks for alerts cleared that should have been escalated, and for alerts escalated that should have been closed. Both errors matter to an examiner, and AI can shift the balance between them.

Some institutions add a testing layer between operations and audit. A head of BSA/AML technology and program assurance at a sponsor bank described building one, so that when auditors or regulators arrive, the team can show the targeted and thematic testing it has already done. The findings that layer looks for are ordinary ones: the MLRO at the payments company said regulators call out missing information, numbers that do not foot, and write-ups that leave out transactions in the alert.

AI calls for tighter guardrails on quality control. A chief BSA officer at a sponsor bank said: "Even if we could automate all the way through SAR submission, just because we can, doesn't mean we should, from a regulatory posture, they're going to want to see us having very tight guardrails on how we QC this."

What AI does not change

Five things in an AML program stay where they are when AI arrives: who is accountable, how the institution judges its risk, who decides to file a SAR, who tests the program, and the referrals people make on their own.

• The BSA officer's accountability. The regulation requires the designation of an individual or individuals responsible for coordinating and monitoring day-to-day compliance. It names a person, not a system, and AI does not move that accountability.

• The risk assessment. FinCEN's April 2026 proposal would fold the risk assessment into the program's internal policies, procedures and controls. It would remain the institution's judgment, informed by better data but not handed to a tool.

• The SAR decision. Deciding that activity is suspicious, and filing, stays with the institution. A director of compliance and BSA at a community bank wanted assurance that the AI would not label activity suspicious at all, because in that director's reading the label would start the filing clock, and the determination belongs to the team.

• Independent testing. Testing stays independent of the people and tools it tests. A BSA officer at a community bank noted that regulators like to see a third party test the program, and that documentation produced by the system under test is not the same thing.

• Referrals people make on their own. Reviewers still spot activity that never alerted. The director of AML compliance mentioned above wanted to keep exactly that: analysts who notice something wrong during other work and refer it for investigation.

A regulator will ask to see where those decisions happen. The same MLRO said the question to prepare for is how the institution exercises oversight: where individual people make the decisions on escalation and SAR filing, and how the institution can demonstrate those decision points. For how those review points are designed, see what human-in-the-loop means in fraud and AML.

How examiners will read an AI-supported AML program

Examiners will read an AI-supported AML program the way they read any program: they will test whether it works and whether the institution can explain it. With AI in the workflow, three questions get sharper.

Does the institution understand its own system?

The community bank director of compliance and BSA mentioned above said, on another call, that a system the team cannot understand will be very difficult to defend in front of regulators, because regulators like knowing the institution understands how a system works. That test applies to every model and agent in the program, including ones a vendor supplies.

Can every outcome be reconstructed?

The same director described the examiner's question directly: "how did the system determine this? And if I don't have it in writing, I won't be able to defend it." Auditors ask the same of manual work. A compliance manager at a consumer lender said that after a BSA/AML audit, the auditors looked at the documented findings and asked where the proof was.

Has it been validated and tested?

The senior sanctions compliance officer mentioned earlier expects every system, AI included, to go through validation under the institution's model risk policy, with extensive testing to show the regulator the bank is comfortable with it. Many teams run AI in parallel with their existing process first; a risk management leader at a sponsor bank described running things in parallel to get regulators comfortable over time. A chief operating officer at a bank said the team is playing it safe until it knows where examiners will draw the boundaries.

Where the rules stand in October 2026

• Model risk. SR 26-2, OCC Bulletin 2026-13 and FDIC FIL-15-2026 have applied since 17 April 2026. Machine-learning monitoring models are in scope, generative and agentic AI are out of scope, and the planned request for information on AI in model risk has not been issued.

• The AML program rule. FinCEN's 10 April 2026 proposal (91 FR 18704) states that FinCEN will not finalise its July 2024 proposal, which the new proposal withdraws and supersedes. For banks, the new proposal sets out four components: risk-based internal policies, procedures and controls (which absorb the risk assessment and ongoing customer due diligence), independent testing, a designated U.S.-located individual responsible for the program, and ongoing training. Comments closed on 9 June 2026, and the rule is not final as of October 2026, so 31 CFR 1020.210 and its five elements still apply.

• Innovation. The December 2018 joint statement on innovative efforts to combat money laundering (SR 18-10, OCC Bulletin 2018-44, FIL-79-2018) encourages banks to consider, evaluate and, where appropriate, responsibly implement innovative approaches. The preamble to FinCEN's 2026 proposal encourages institutions to evaluate whether new technology might help combat financial crime more effectively, and names machine learning and generative AI among the examples. A proposal's preamble is not a rule, and neither document approves any particular tool.

Where AI-supported programs fall short

• No written record of what the AI system saw and why it recommended what it did.

• Model, threshold or scenario changes that reach production before they are approved.

• Treating review of each AI output as the program's quality assurance, and dropping second-line testing.

• Treating a vendor's own testing as the program's independent testing.

• Letting an AI system label activity suspicious, which takes a judgment that belongs to the institution.

• Describing an agent as covered by SR 26-2 when the guidance places agentic AI outside its scope.

Frequently asked questions

Can AI file a SAR?

The decision to file a SAR, and the filing itself, stay with the institution. AI can help assemble the case and draft the narrative, and people review, sign and submit the report. Our piece on SAR narratives covers how drafting fits with human sign-off.

Does using AI change the BSA officer's responsibilities?

No. The BSA officer remains the individual responsible for coordinating and monitoring day-to-day compliance, and that includes the AI systems in the program. What changes is the scope of oversight: models, agents and the controls around them become part of the program the officer answers for.

Does model risk management guidance apply to AI in AML?

It applies to machine-learning models used in AML, such as transaction monitoring and screening models. The revised interagency guidance issued on 17 April 2026 (SR 26-2, OCC Bulletin 2026-13, FDIC FIL-15-2026) replaced SR 11-7 and the 2021 BSA/AML statement, and it places generative and agentic AI outside its scope. Institutions still need validation and controls for those systems, and the agencies' planned request for information on AI in model risk has not been issued as of October 2026.

Is FinCEN's AML program rule changing?

It may. FinCEN proposed a revised AML/CFT program rule on 10 April 2026, replacing its withdrawn 2024 proposal, which would give banks four program components instead of today's five elements. Comments closed on 9 June 2026 and the rule is not final as of October 2026, so 31 CFR 1020.210 still applies.

What audit trail should an AI-supported AML program keep?

Enough to reconstruct each outcome: what data the AI system saw, what it recommended and why, which model or version produced it, who reviewed it and what that person decided. Changes to models and thresholds need their own record of approval. An examiner who asks how the system reached a conclusion should get a written answer.

Where Oscilar fits

Oscilar is built so a person confirms each recommendation. The platform keeps a full audit trail per decision with stored reasoning, and human-in-the-loop by design, so an institution can show an examiner where its people decided. The platform is designed to fit the interagency model risk management guidance. That guidance places generative and agentic AI out of scope, so any AI agent in the program is governed through the institution's own controls and review points rather than through that guidance.

For what an AI agent in risk work is, and what it is not, see what an autonomous AI agent is in risk work. If you are assessing vendors, our guide to evaluating an AI-powered AML platform sets out the questions to ask.

For a practical view of how AI is deployed across AML and FRAML functions today, read our AI in AML playbook, developed in collaboration with FS Vector.