The 2018 AML innovation pledge protected the pilot, not the switch

On December 3, 2018, five agencies promised that an AI pilot finding what old monitoring missed would not by itself prove the old system deficient. That cleared one barrier. Deciding when to switch, validating the model and reporting what it finds stayed with the bank.

Looking Back revisits past weeks one at a time, up to the most recent, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.

Two-minute summaryThe dek and the first sentence of each section, quoted from the article

On December 3, 2018, five agencies promised that an AI pilot finding what old monitoring missed would not by itself prove the old system deficient. That cleared one barrier. Deciding when to switch, validating the model and reporting what it finds stayed with the bank.

  1. The week On Monday, December 3, 2018, five federal agencies issued a joint statement on innovative efforts to combat money laundering and terrorist financing.
  2. The discovery problem: what the statement solved The fear the agencies addressed had a clear mechanism.
  3. The switch problem: deciding when a pilot becomes the program The statement told bank management to evaluate prudently whether, and at what point, an innovative approach was developed enough to replace or augment existing processes.
  4. The validation problem: proving the model works The third problem is evidence.
  5. The reporting problem: what to do with what the model finds The fourth problem begins when the new system works.
  6. Objections: a statement without teeth, or the wrong barrier One objection is that the 2018 statement was only a statement.
  7. With hindsight The validation question came first.

The week

On Monday, December 3, 2018, five federal agencies issued a joint statement on innovative efforts to combat money laundering and terrorist financing. The signers were the Board of Governors of the Federal Reserve System, the Federal Deposit Insurance Corporation and the Financial Crimes Enforcement Network (FinCEN). The National Credit Union Administration and the Office of the Comptroller of the Currency joined them. The Federal Reserve said the statement came out of a working group those agencies had formed with the Treasury’s Office of Terrorism and Financial Intelligence. The group aimed to improve the effectiveness and efficiency of the Bank Secrecy Act and anti-money laundering (BSA/AML) regime.

The statement was short, three pages in the version the Federal Reserve posted. It encouraged banks to consider, evaluate and, where appropriate, responsibly implement innovative approaches to meet their BSA/AML obligations. A footnote applied the term “bank” as the Bank Secrecy Act defines it, which includes savings associations, credit unions and the U.S. offices of foreign banks. The agencies said innovation could augment risk identification, transaction monitoring and suspicious activity reporting. They noted that some banks were building internal financial intelligence units, and that some were experimenting with artificial intelligence and digital identity technologies.

Two commitments did most of the work. First, the agencies said they would not penalize or criticize banks that maintained effective programs but chose not to innovate, and that they would not advocate any particular method or technology. Second, they addressed pilots. Pilot programs, they said, should not in themselves bring supervisory criticism, even if they failed. Pilots that exposed gaps in a compliance program would not necessarily lead to supervisory action. If a bank tested an AI-based transaction monitoring system and found suspicious activity its existing processes had missed, the agencies would not automatically assume those processes were deficient. They would assess the adequacy of the existing monitoring processes independently of the pilot’s results. Implementing innovative approaches would not, they said, result in additional regulatory expectations.

The statement also set limits. Banks had to keep meeting their BSA/AML obligations and remain safe and sound while they experimented. Management should prudently evaluate whether, and at what point, an innovative approach was developed enough to replace or augment existing processes. That evaluation should weigh information security, third-party risk management, and compliance with other laws, such as those on customer notifications and privacy. Banks should discuss their evaluations with their regulators.

On engagement, FinCEN said it would consider requests for exceptive relief under its regulations to allow testing of new technologies, provided banks kept their programs effective. It announced an innovation initiative with dedicated times for financial institutions and technology firms to meet with it. The other agencies said they had established, or would establish, offices to support responsible innovation. The agencies said they were exploring further ways to encourage innovation, including through FinCEN’s Bank Secrecy Act Advisory Group. They invited industry feedback on how they could support innovation through explanations of, or updates to, supervisory processes, regulations and guidance. FinCEN’s release of the same day quoted the Treasury’s Under Secretary for Terrorism and Financial Intelligence as highly encouraged by recent steps financial institutions had taken to identify and report illicit financial activity related to North Korea, Iran and other priorities.

The December 2018 joint statement can be read as regulators blessing artificial intelligence for anti-money laundering. Its text was narrower and more useful than that. It addressed one specific fear: that a bank testing better monitoring would, by finding more, prove its existing monitoring inadequate and invite criticism for the past. The agencies took that fear off the table. They did not take away the other obstacles between a promising pilot and a production system.

Those obstacles are the subject of this article, which divides the path from pilot to production into four problems. The first is the discovery problem: what happens when a pilot finds what the old system missed. The statement answered it. The second is the switch problem: deciding when a new approach is good enough to replace or supplement the old one, and running both in the meantime. The third is the validation problem: proving to the bank’s own risk managers and its examiners that the new model works as intended. The fourth is the reporting problem: handling, within the Bank Secrecy Act’s deadlines, the suspicious activity the new system surfaces. A final section takes the main objections.

The four problems are separate and sequential. Discovery concerns supervisory treatment of the past. The switch concerns operating decisions in the present. Validation concerns evidence about the model. Reporting concerns obligations triggered by what the model finds. A bank can resolve one and stall on another. The statement removed the first obstacle. The later history, through the 2026 proposals and revised model risk guidance, is largely the story of the other three.

The discovery problem: what the statement solved

The fear the agencies addressed had a clear mechanism. A bank runs a new monitoring approach alongside its existing one. The new approach flags customers or transactions the old one did not. Each flag that turns out to be genuinely suspicious is evidence that the old system missed something. A bank in that position could reasonably worry that an examiner would treat the pilot’s results as proof of a deficient program and require remediation of everything the old system might have missed.

The statement spoke to that mechanism directly. If a pilot of an AI-based transaction monitoring system found suspicious activity that existing processes would not have caught, the agencies said they would not automatically assume the existing processes were deficient. They would assess those processes independently of the pilot’s results. Pilots that exposed gaps would not necessarily lead to supervisory action, and unsuccessful pilots would not in themselves draw criticism.

Why the promise mattered to boards

The promise mattered most at the level where investment decisions are made. A compliance officer who wants to test a new monitoring method needs budget, data access and the board’s tolerance for risk. If the board believes the test could generate evidence against the bank, it has a rational reason to say no. The statement gave the compliance officer an answer: the agencies had said in writing that the test itself would not be held against the bank.

Boards may also look for signs that regulators will not penalize inaction. The statement covered that too. Banks that kept effective programs but chose not to innovate would not be penalized or criticized. For a small institution with limited resources, that sentence meant it could watch larger peers experiment without feeling pressure to follow before it was ready.

What the promise did not cover

The wording was careful. Pilots would not “necessarily” lead to supervisory action, and the agencies would not “automatically” assume deficiency. Those qualifiers left room for examiners to act if a pilot revealed a serious, systemic gap. The statement also kept the core obligation intact: banks had to continue to meet their BSA/AML obligations while experimenting.

A careful reader would take two operating rules from that wording. The first is that the pilot itself is protected, but what the bank does after learning of a gap is not. If a pilot reveals that the old system misses a whole class of activity, the bank still has to decide what to do about that class. The second is that the existing program must stand on its own. A bank cannot point to the pilot as a substitute for fixing known weaknesses in the production system.

Designing a pilot to use the protection

A bank that wants the benefit of the statement should design its pilot so the protection clearly applies. That means documenting before the pilot starts that it is a test, what it is testing, what data it uses and how long it will run. It means keeping the existing program fully operational. And it means planning in advance what the bank will do with the pilot’s findings, both the individual alerts and any pattern of gaps.

The statement invited early engagement. The agencies said they were open to discussing pilot programs with bank management, and that early engagement could help them understand the approach and discuss expectations. A bank that briefs its examiners before a pilot has a record that the test was planned as a test. That record is more persuasive than an explanation offered after a difficult finding.

Exceptive relief and innovation offices

The statement offered two further tools that bear on the discovery problem. FinCEN said it would consider requests for exceptive relief under its regulations to facilitate testing of new technologies, provided banks maintained the overall effectiveness of their programs. That gave a bank a formal route to ask for room when a test could not be run within existing requirements.

The other agencies said they had or would establish projects or offices to support responsible innovation. Those offices could serve as central points of contact, while banks continued to follow existing protocols for communicating with their regulators. For a bank, the practical meaning was that the innovation office supplements the examination relationship and does not replace it. A pilot discussed with an innovation office should also be discussed with the bank’s examination team, so that the people who will later assess the program know the test was planned.

Who benefits most

The discovery protection mattered most to institutions with enough data and staff to run a meaningful pilot. A large bank with its own data science team could test a new model across millions of transactions. A credit union or community bank might instead depend on its core processor or a monitoring vendor to offer something new. For those institutions, the protection applied to pilots their vendors ran with them, which put a premium on vendor contracts that allowed testing and shared the results with the institution.

The switch problem: deciding when a pilot becomes the program

The statement told bank management to evaluate prudently whether, and at what point, an innovative approach was developed enough to replace or augment existing processes. That sentence describes the hardest operating decision in the whole process, and the statement gave no test for it. The decision stayed with the bank.

The difficulty is that two monitoring systems can disagree. A new model will flag some activity the old rules miss and miss some activity the old rules flag. Switching means accepting the second set of misses in exchange for the first set of finds. Augmenting means running both, with the cost of investigating alerts from each.

Replace or augment

The choice between replacing and augmenting has different costs. Augmenting is safer in one sense: nothing the old system caught is lost. But it increases the alert volume and the staff needed to work it, and it can leave the bank with two sets of logic that nobody fully understands together. Replacing is cheaper to run but requires confidence that the new system’s misses are acceptable.

A bank deciding between them needs evidence on overlap. It should know which alerts from the old system led to suspicious activity reports, and whether the new system would have caught them. It should know which of the new system’s alerts led to reports, and what kinds of activity they involved. The answers show whether the new approach covers the risks that matter. They should be measured during the parallel run, not estimated afterward.

The cost of the parallel run

Running two systems in parallel costs money and staff time. Every alert from the new system needs a human to decide whether it is worth investigating, at least until the bank trusts the model. If the new system generates many alerts, the investigation team can be overwhelmed. If the bank does not staff up, alerts age, and an old alert that turns out to be genuine raises timing questions under the suspicious activity reporting rules discussed below.

That cost argues for scoping. A bank can pilot the new approach on one product, one customer segment or one typology, rather than across the whole bank. A narrow pilot produces less evidence but fewer alerts, and it lets the bank learn how to staff and govern the new system before scaling it.

A hypothetical credit union pilot

Consider, as an illustration, a mid-sized credit union whose monitoring vendor offers a new machine learning module. The vendor proposes running it on the credit union’s data for six months alongside the existing rules. The credit union’s BSA officer sees the appeal: fewer false positives, and perhaps detection of patterns the rules miss.

The 2018 statement lets the BSA officer tell the board that the pilot itself will not be held against the credit union. The harder questions come next. The team of four investigators cannot work double the alerts, so the pilot is scoped to member business accounts. The vendor agrees to provide documentation and test results, written into an amendment. The privacy officer reviews whether new data uses require member notices. The credit union briefs its examiner before the pilot starts and records the briefing.

Six months later, the module has flagged a few cases the rules missed and missed a few the rules caught. The credit union now faces the switch decision with evidence rather than a sales presentation. Whether it replaces, augments or declines, it can explain why. That is the outcome the statement made possible, and the work that produced it was the credit union’s.

Thresholds and tuning

Much of the practical difference between monitoring approaches can lie in thresholds. A model that ranks risk still needs a cutoff above which a case is opened. Set too low, and investigators drown. Set too high, and the bank misses activity. The old rules-based systems faced the same choice, but their logic was easier to explain. A model’s cutoff produces outcomes that depend on how the model weighs many factors at once.

Tuning therefore needs governance. The bank should document how thresholds are chosen, who approves changes, and how the effect of each change on alert volumes and reporting outcomes is measured. Without that record, a bank cannot explain to an examiner why it stopped generating alerts on a type of activity, which is a question that can follow a sharp drop in alert volume.

Other laws in the switch decision

The statement named factors beyond BSA/AML that management should weigh: information security, third-party risk management, and compliance with other laws such as those on customer notifications and privacy. Each can block a switch that looks good on detection metrics alone.

Information security matters because a new model can need data from many systems, concentrated in one place. Third-party risk matters because some new approaches come from vendors whose models the bank did not build. Customer notification and privacy laws matter when the new approach uses data in ways the bank has not told customers about, or when digital identity methods collect new kinds of personal information. A switch plan should show how each was addressed, with the owner of each sign-off named.

What the evidence pack for a switch contains

When a bank decides to switch or augment, it should be able to show the decision was prudent, in the statement’s word. A sound evidence pack would include the following.

  • The pilot design, scope and duration, recorded before the pilot began.
  • Overlap analysis between the old and new systems, measured on real alerts and reports.
  • The threshold-setting method and the expected change in alert volumes and staffing.
  • Sign-offs from information security, third-party risk and privacy owners.
  • A record of what was discussed with examiners and when.

This pack is not a regulatory checklist. It is the record a bank would want if an examiner later asked why it changed systems, and it costs little to assemble while the work is happening.

The validation problem: proving the model works

The third problem is evidence. A bank that adopts a new monitoring model has to show, to its own risk managers and to examiners, that the model does what the bank says it does. The 2018 statement did not address validation. It came back to the surface in April 2021 and remained unsettled into 2026.

In 2021, the OCC, the Federal Reserve and the FDIC, in consultation with the NCUA and FinCEN, issued a statement on the subject. It clarified that the principles of the interagency model risk management guidance were appropriate considerations for BSA/AML systems. At the same time, the five agencies published a request for information. They asked how far those principles supported compliance with BSA/AML and sanctions requirements, and whether more explanation could increase transparency, effectiveness or efficiency.

Why model risk guidance was a poor fit, in some views

The 2021 request showed where the tension lay. It asked whether banks’ model risk frameworks included testing and validation that went further than the independent testing the Bank Secrecy Act already requires. It asked whether banks used outside parties to validate BSA/AML and sanctions systems, and whether their model risk frameworks applied to all models, including BSA/AML and sanctions models. It asked about the particular challenges of internally developed and third-party systems.

Those questions reflect a practical problem. Transaction monitoring systems mix rules, scenarios, thresholds and sometimes statistical models. A validation framework has to decide which of those components it treats as models, and how deeply it tests each one. If every element is treated as a model, validation can become extensive and slow. A bank that wanted to change a threshold quickly could find itself waiting for a validation cycle. That delay can discourage exactly the experimentation the 2018 statement encouraged.

What validation of a monitoring model should cover

Whatever the formal framework, validation of an AML model has to establish five things. The model uses the data the bank thinks it uses, and that data is complete. Its output corresponds to real risk, as shown by which alerts lead to reports. Its performance is stable over time and across customer segments. Its thresholds are documented and justified. And the bank can explain, for a particular case, why the model flagged it or did not.

The last question matters for investigators as well as examiners. An investigator who receives a model-generated alert needs to know why it was generated to work the case efficiently. A model that produces scores without reasons leaves investigators to reconstruct the logic. Explanation is therefore an operating requirement as well as a governance one.

Data quality comes before model quality

Validation also exposes the data underneath a model. Transaction monitoring depends on customer records, account relationships, transaction descriptions and counterparty details drawn from several systems. Missing or inconsistent fields produce blind spots that no model can fix. A new model can make those blind spots more visible, because it uses data the old rules ignored.

Banks should treat data lineage for AML as part of validation. That means knowing which source systems feed monitoring, how often, with what transformations and with what checks for completeness. A model validated on clean test data and deployed on incomplete production feeds can underperform in ways that are hard to detect. The validation record should show that production data was tested as well as test data.

The 2026 turn

In April 2026, the federal banking supervisors and FinCEN moved on this question from two directions. On April 17, 2026, the Federal Reserve issued SR 26-2, announcing revised model risk management guidance from the Federal Reserve, the OCC and the FDIC. It superseded both the 2011 model risk guidance and the 2021 BSA/AML model risk statement. The letter said the revision emphasized a risk-based approach tailored to each organization’s model risk profile and to the size and complexity of its operations. It said the letter was expected to be most relevant to banking organizations with over $30 billion in total assets regulated by the Federal Reserve.

Earlier that month, in its April 10, 2026 proposal on AML/CFT programs, FinCEN discussed comments it had received on the 2021 request. Commenters had worried that supervisors might expect banks to apply the model risk guidance to AML/CFT controls. FinCEN said that, while it had not issued or been party to that guidance, it shared certain of those concerns: such models may be overly burdensome and ill-fitted to address illicit finance risks. FinCEN invited comment on that position and said it intended to work with the banking agencies on those concerns.

What a bank should do while the framework settles

For a bank, the practical response is to validate AML models in proportion to their role and risk, and to document the reasoning. A model that decides which alerts investigators never see warrants deeper validation than one that reorders a queue. A vendor model warrants testing on the bank’s own data, because performance on the vendor’s reference data says little about the bank’s customers. A threshold change warrants measurement of its effect on outcomes, even if it does not warrant a full revalidation.

That approach fits the risk-based emphasis of the 2026 guidance and does not depend on how FinCEN’s concerns are resolved. It also produces the evidence an examiner will ask for under any framework: what the model does, how the bank knows, and what the bank does when performance changes.

Vendor models and the limits of access

Where an institution uses a monitoring system built by a vendor, validation depends on what the vendor will disclose. A bank that cannot see how a vendor model works can still test its outputs on the bank’s data, compare alerts to outcomes, and monitor stability over time. It should also make sure its contract gives it the documentation, test results and change notices it needs. The 2021 request noted that the decision to use an outside party for validation was the bank’s own, made under its third-party and model risk management requirements.

The reporting problem: what to do with what the model finds

The fourth problem begins when the new system works. A pilot that finds suspicious activity the old system missed has found suspicious activity. The statement protected the bank from automatic criticism of its old system. It did not suspend the bank’s obligation to report.

For banks, FinCEN’s rule requires a suspicious activity report no later than 30 calendar days after the date the bank initially detects facts that may constitute a basis for filing. If no suspect has been identified on that date, the bank may take an additional 30 calendar days to identify one. Reporting may not be delayed more than 60 calendar days after initial detection. Banks must keep a copy of each report and its supporting documentation for five years.

When does a pilot detect

A pilot raises a timing question the regulations do not answer in terms. When a model flags a transaction during a test, one reading is that the bank has then detected facts that may constitute a basis for filing. Another is that detection occurs when an investigator reviews the alert. The answer affects how quickly pilot alerts must be worked. Organizations should check with counsel on the right reading for their circumstances.

The safe operating assumption is that alerts from a pilot need to be reviewed on a timetable consistent with the bank’s ordinary alert-handling process. A pilot that generates alerts nobody reviews for months creates a backlog that may be hard to defend. That is one more reason to scope pilots so the investigation team can keep up.

Lookbacks

A new model can also be run on historical data, which is one way to demonstrate its value. If the historical run flags old activity that looks genuinely suspicious, the bank faces a decision about whether those findings require reports. A historical analysis designed as a test of the model is still an analysis of real customers’ transactions.

Banks should decide before any historical run how they will treat its findings. The decision should be made with counsel, recorded, and communicated to the team running the analysis. Leaving it until after the results arrive invites inconsistent handling. A mix of filed and unfiled findings is harder to explain than a clear decision to review all of them, or a clear rationale for the approach taken.

Investigation capacity

In the end, the reporting problem is a capacity problem. Better detection produces more cases that need human judgment. If the investigation team does not grow, or does not get better tools, more cases mean slower handling, and slower handling pressures the 30-day clock.

Innovation can help on this side too. The statement mentioned financial intelligence units that identify complex illicit finance threats, and the same technologies that generate alerts can help investigators. They can assemble case files, link related accounts and summarize activity. A bank planning a new detection model should plan its effect on investigation workload at the same time, and fund both.

Quality of reports as well as quantity

More alerts do not automatically mean better reports. The purpose of reporting is to give law enforcement useful information. A model that finds genuinely new patterns can improve the reports a bank files. A model that simply generates more marginal cases can increase filings without adding value. The bank’s measures of success should include outcomes that matter to the reporting purpose, such as the share of alerts that lead to reports and the quality of the narratives, alongside detection rates.

FinCEN’s April 2026 proposal pointed in this direction. Among the factors its Director would consider in deciding whether to take an enforcement or significant supervisory action against a bank, it included conducting proactive analytics or other innovative activities producing demonstrable outputs evincing the program’s effectiveness. Its examples included effective use of artificial intelligence, federated learning or other advanced monitoring tools. If adopted in that form, the emphasis would fall on what a program produces as well as on how it is built.

Privacy and customer notification

The statement told banks to consider compliance with laws on customer notifications and privacy when evaluating innovative approaches. That instruction bears on reporting as much as on the switch. New approaches may draw on data the bank did not previously use for monitoring, such as device and digital identity signals, or may share data with a vendor’s platform. Each use needs a basis under the bank’s privacy obligations and its own notices to customers.

Suspicious activity reporting adds a constraint of its own. FinCEN’s rule for banks makes a report, and any information that would reveal its existence, confidential. It bars disclosure by any director, officer, employee or agent of the bank except as the rule allows. A vendor platform that processes alerts and case files therefore handles some of the most sensitive information a bank holds. The contract, access controls and logging for that platform should reflect that sensitivity. Counsel should confirm how confidentiality obligations apply to the vendor’s staff and systems.

Credit unions and community banks

The reporting problem looks different for smaller institutions. A community bank or credit union with a small BSA team may lack the capacity to absorb a surge of new alerts. For such an institution, the right innovation may be one that improves the quality of fewer alerts rather than one that finds more. Its pilot should be judged by whether investigators spend their limited time on better cases, which is a measure a small team can track easily.

Objections: a statement without teeth, or the wrong barrier

One objection is that the 2018 statement was only a statement. It created no safe harbor in regulation, and its qualifiers left examiners free to act. On this view, a cautious bank had little reason to rely on it.

The objection is correct about legal force and less so about practical effect. A written commitment from five federal banking and anti-money laundering authorities can shape how examination teams approach a pilot, and it gave compliance officers something concrete to show their boards. Its qualifiers were also reasonable. A promise never to act on a pilot’s discoveries would have been one no supervisor could keep.

The objection that the real barrier was cost

A second objection holds that fear of criticism was never the main barrier. The main barriers were cost, data quality and the scarcity of people who understand both machine learning and money laundering. On this view, the statement addressed a secondary concern.

There is truth in this, and it is consistent with the argument here. The statement removed one barrier, the discovery problem, and left the switch, validation and reporting problems in place. Cost, data and skills are the substance of those three. The agencies could promise not to penalize a pilot. They could not fund the parallel run, clean the bank’s data or hire its investigators.

The distinction matters for how a bank reads later encouragement. When the 2026 proposals again urged institutions to evaluate machine learning, generative AI and digital identity, they repeated the 2018 message on the discovery side. A bank deciding whether to act on that encouragement should look first at its switch, validation and reporting capacity. If those are weak, a pilot will produce findings the bank cannot use, however welcome the regulators make the attempt.

The objection that model risk rules blocked innovation

A third objection is that supervisory expectations on model validation, applied after 2021, undercut the 2018 encouragement. FinCEN’s own April 2026 comments lend this some support, since it said it shared commenters’ concerns that such models may be overly burdensome and ill-fitted to address illicit finance risks.

That critique is about how validation was applied, not about whether validation is needed. A bank still needs to know whether its model works. The 2026 revised guidance’s emphasis on a risk-based, tailored approach, and FinCEN’s stated intention to work with the agencies, suggest the debate is over proportion. A bank that validates in proportion to risk, and records its reasoning, is positioned for either outcome.

What we would tell a chief compliance officer today

Treat the path from pilot to production as four separate problems with four owners. The pilot design and examiner engagement belong to the BSA officer. The switch decision belongs to a committee with business, compliance, security, privacy and third-party risk represented. Validation belongs to model risk or an equivalent independent function, scaled to the model’s role. Reporting capacity belongs to the head of investigations, with a budget tied to expected alert volumes.

A bank that assigns those owners before it begins a pilot will move faster than one that discovers each problem in sequence. The 2018 statement made it safer to start. The work of finishing has always been the bank’s.

What would change the argument

The argument would be weaker if supervisors had later treated a successful pilot as automatic evidence of past deficiency, or if the switch, validation and reporting problems had been resolved by rule. We found no public record of either as of November 2026. The 2026 proposals encouraged innovation in terms close to the 2018 statement’s. The revised model risk guidance emphasized proportion. The reporting rules were unchanged. On the public record, the barrier the statement removed stayed removed, and the other three remained the bank’s to solve.

With hindsight

The validation question came first. On April 12, 2021, the OCC, the Federal Reserve, the FDIC, the NCUA and FinCEN asked for public comment on how far model risk management principles supported BSA/AML and sanctions compliance. Three of the agencies issued a concurrent statement that those principles were appropriate considerations for BSA/AML systems. The request asked how far banks’ model risk frameworks went beyond the independent testing the Bank Secrecy Act already requires. It also asked whether banks used outside parties to validate their systems, and what challenges internal and vendor systems raised. Comments were due June 11, 2021.

In April 2026, the ground shifted. On April 10, 2026, FinCEN proposed a rule to fundamentally reform AML/CFT program requirements and to implement provisions of the Anti-Money Laundering Act of 2020. It said it did not intend to finalize its 2024 program proposal, which should be considered withdrawn and superseded, and noted that the 2024 proposal had drawn 86 comments. The new proposal cited the 2018 joint statement and encouraged institutions to evaluate machine learning, generative AI, digital identity, blockchain monitoring and analytics, and application programming interfaces. It also said FinCEN shared commenters’ concerns that such models may be ill-fitted to address illicit finance risks. The OCC, FDIC and NCUA published a companion proposal the same day, quoting the AML Act’s purpose of encouraging technological innovation. Comments on FinCEN’s proposal were due June 9, 2026. The proposal would also create a two-pronged framework, under which a program would be deemed effective if the institution establishes and maintains it, and maintaining it means implementing the properly established program in all material respects. It proposed enhancing FinCEN’s role in AML/CFT supervision and enforcement, in coordination with the federal banking regulators. Among the factors its Director would consider before an enforcement or significant supervisory action against a bank were innovative activities producing demonstrable outputs, including effective use of artificial intelligence.

On April 17, 2026, the Federal Reserve issued SR 26-2. Issued jointly with the OCC and FDIC, the revised guidance replaced the 2011 model risk guidance and the 2021 BSA/AML model risk statement. It emphasized a risk-based approach tailored to each organization’s model risk profile and to the size and complexity of its operations. The letter said the revision reflected fifteen years of supervisory experience and industry feedback, and advances in modeling practice.

As checked on November 14, 2026, neither April 2026 program proposal had been finalized in the Federal Register. The suspicious activity reporting deadlines for banks remained as they were in 2018: 30 calendar days from initial detection, extendable to 60 when no suspect has been identified. For banks, the practical result is a split timetable. Supervisory guidance on validating models changed in 2026, while the rule on what counts as an effective program, and how examiners will weigh innovation within it, remained open.

What held up was the statement’s central promise and its encouragement, which the 2026 proposals echoed. What did not hold up was the assumption that clearing the discovery problem would be enough. The validation question took until 2026 to reach revised guidance, and the program rule that might settle how supervisors judge effectiveness was still a proposal.

Spotted an error? Tell us and we will check it and correct the article where needed.

  • anti-money laundering
  • Bank Secrecy Act
  • transaction monitoring
  • model risk management
  • FinCEN