AI Governance: A Practical Priority for Community Bank Compliance
Artificial intelligence is already embedded in many community bank operations. Fraud monitoring systems, credit decisioning tools, customer chatbots, and vendor-provided analytics often rely on AI or algorithmic processes and may have been implemented before the bank understood that AI was involved. Employees may use these tools without realizing that the oversight expectations applied to traditional models and third-party relationships may also apply here. Community banks should adopt a coherent, risk-based AI governance approach as part of ordinary risk management rather than wait for a formal regulatory mandate.
The United States does not yet have a single comprehensive federal AI statute as of the date of this article, but the guidance landscape has matured quickly over the last few years, with recent efforts in the banking industry providing clearer direction on governance needs. Several documents already provide useful expectations and frameworks for responsible use of AI in financial services that offer a workable foundation for community banks to establish and build upon as frameworks and laws are formalized and the industry learns from early adoption and oversight challenges.
Current U.S. AI Governance Guidance Relevant to Banks
NIST AI Risk Management Framework (AI RMF 1.0). The NIST AI Risk Management Framework remains one of the federal government’s primary voluntary frameworks for trustworthy and responsible AI. Published on January 26, 2023, it organizes risk-management activities around four functions: Govern, Map, Measure, and Manage. Its principles concerning fairness, transparency, accountability, privacy, reliability, and safety align well with existing banking risk-management obligations. NIST is currently revising AI RMF 1.0, but it remains a useful baseline for organizations managing AI risk.
Financial Services AI Risk Management Framework (FS AI RMF). In February 2026, the U.S. Department of the Treasury announced the Financial Services AI Risk Management Framework along with a shared AI Lexicon for the financial sector. The FS AI RMF was developed through a public-private collaboration led by the Cyber Risk Institute in coordination with the Financial Services Sector Coordinating Council, with participation from more than 100 financial institutions and input from government agencies, including NIST. The framework adapts the NIST structure specifically for financial services and addresses AI risk across areas such as credit, fraud, AML, operations, customer interactions, and vendor-supplied systems. It includes practical tools such as an AI Adoption Stage Questionnaire and a Risk and Control Matrix containing 230 control objectives. While voluntary, it is designed to scale to financial institutions of different sizes and provides a useful reference point for governance, audit, and examination readiness.
Revised Interagency Model Risk Management Guidance (SR 26-2 / OCC Bulletin 2026-13). Issued on April 17, 2026, this guidance updates the longstanding model risk framework and replaces prior interagency model risk guidance. It reaffirms the importance of model inventories, validation, performance monitoring, documentation, governance, and effective challenge while emphasizing a risk-based approach tailored to an institution’s model risk profile, size, and complexity. The Federal Reserve states that the guidance is expected to be most relevant to banking organizations with more than $30 billion in total assets. Although many community banks fall below that threshold, the risk-management concepts may still provide useful guidance where appropriate. Generative and agentic AI models are outside the formal scope of the guidance because those technologies are novel and rapidly evolving. The agencies nevertheless state that a banking organization’s risk-management and governance practices should guide appropriate controls for tools outside the guidance. Traditional statistical and quantitative models and non-generative, non-agentic AI models remain within its scope.
CFPB expectations under ECOA/Regulation B, UDAAP, and related guidance. CFPB guidance issued in 2022 and 2023 emphasized that the use of AI or complex algorithms does not relieve a creditor of its obligation under the Equal Credit Opportunity Act (ECOA) and Regulation B to provide specific and accurate adverse-action reasons. Creditors cannot rely on a model being too complex or opaque to explain the actual principal reasons for an adverse action. More broadly, existing fair-lending and consumer-protection laws continue to apply when AI is used in credit decisioning, underwriting, marketing, product selection, or other consumer-facing activities.
NAIC AI Model Bulletin. Although written for insurers, the National Association of Insurance Commissioners’ Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023, can provide a useful reference for banks. It emphasizes governance, risk management, fairness, accountability, transparency, and controls surrounding AI systems. Its approach also reinforces the importance of adapting third-party risk and data-governance processes as AI becomes more widely used.
American Bar Association Formal Opinion 512 (July 2024). Issued for attorneys, the opinion highlights confidentiality and other professional-responsibility risks associated with the use of generative AI. The practical lesson for banks is straightforward: nonpublic personal information, proprietary data, and confidential bank materials should not be entered into external AI tools unless the bank has approved the tool and established appropriate data-sharing, confidentiality, and security controls.
American Bankers Association resources. In June 2026, the American Bankers Association published an AI Policy Template designed to help banks establish an AI policy. ABA also makes available The AI Framework for Community Banks, which emphasizes practical governance, oversight, and integration of AI considerations into existing risk, compliance, cybersecurity, and operational programs. Community banks can use these resources as a starting point for policy and governance development.
Independent Community Bankers of America (ICBA) Community Banker AI Security Readiness Guide. ICBA released its Community Banker AI Security Readiness Guide on June 3, 2026. Developed with community bankers serving on ICBA’s AI Task Force, the resource highlights the changing AI security environment, including AI-enabled fraud and other emerging threats. It also addresses third-party due diligence, vendor dependence, contract language, incident response planning, board oversight, and cyber insurance considerations.
Why Community Banks Should Act Now
AI is appearing in workflows faster than formal governance structures are being built. Employees may use AI-enabled features without realizing the tools require oversight. Vendors increasingly embed AI in core systems, digital banking platforms, fraud tools, and analytics products, so banks often use AI indirectly. Without a basic inventory, these systems can remain invisible and unmanaged.
The risks are concrete and familiar:
- Consumer harm from biased or insufficiently transparent credit or underwriting decisions.
- UDAAP exposure when AI influences product offers, marketing, or customer interactions.
- Data leakage if nonpublic personal information or proprietary information is entered into external AI tools.
- Model performance drift when models are not monitored against changing economic conditions and borrower behavior.
- Third-party risk when vendors use AI without adequate disclosure or documentation.
- Gaps in fairness or explainability arising from incomplete data or unexamined assumptions.
The difference is that AI can amplify these risks quickly and across multiple business lines at once.
A Practical Governance Approach for Community Banks
Community banks do not need a large, specialized team to begin. Most already maintain model risk, third-party risk, data governance, information security, and compliance programs. The task is to extend those programs deliberately to AI and algorithmic tools.
Start with a complete inventory. Identify every AI system, machine-learning model, algorithmic decisioning tool, and vendor product that uses AI or automated decision logic. Include tools used in credit, fraud, AML, operations, marketing, and customer service, whether developed internally or supplied by a vendor.
Once the inventory exists, apply existing controls:
- Assess each system for fairness, transparency, explainability, and consumer impact, especially where credit, underwriting, or product decisions are involved.
- Prohibit the entry of nonpublic personal information, proprietary data, or other sensitive bank information into external AI tools unless the arrangement has been reviewed and approved. Reinforce the rule through regular staff training.
- Apply model risk-management principles where appropriate: validation, performance monitoring, documentation of assumptions, and effective challenge. For generative tools outside formal model guidance, document the purpose, data-handling rules, human-review requirements, and escalation paths.
- Strengthen third-party risk management so that vendors disclose AI use and, where relevant and reasonably available, provide information regarding data sources, decisioning processes, testing, controls, and documentation sufficient to meet the bank’s risk-management standards.
- Establish a written AI policy or update existing policies to address inventory, human oversight, data protections, vendor expectations, and board or senior management reporting. ABA’s policy template and related community-bank materials offer a practical starting point.
- Human oversight should remain an important control for material or higher-risk uses where appropriate. Document where that review occurs and who is responsible. Lifecycle monitoring, including checking that systems continue to perform as expected, should be built into existing risk and compliance calendars.
What to Do Next
Community banks should begin laying the groundwork now. Start by building an inventory of every AI-enabled tool already in use, including those embedded in vendor platforms. Clarify internal data-handling expectations so staff understand what information may and may not be shared with external systems. Ask key vendors to disclose how AI is incorporated into their products and what documentation, testing results, or model explanations they can provide. Use these insights to draft or update a concise AI policy that matches the bank’s size, complexity, and risk profile.
For significant or novel AI uses, banks should also consider whether discussion with appropriate supervisory contacts is warranted, particularly when the use could materially affect consumer compliance, model risk, cybersecurity, or other significant risk areas.
Banks that take these steps now will be positioned to adopt AI with greater confidence through the traditional risk-based approach of protecting consumers, safeguarding sensitive data, and meeting evolving regulatory expectations. This work is not a departure from traditional risk management; it is an extension of the discipline the industry already knows well. Starting early simply reduces the scramble later and helps ensure the institution is ready for the rapid evolution ahead.
Dana Paull, CRCM
Virtual Compliance Officer
If It Walks Like a Rule…: FinCEN’s Minnesota GTO After Novedades
“There’s an old saying in Tennessee – I know it’s in Texas, probably in Tennessee – that says, fool me once, shame on – shame on you. Fool me – you can’t get fooled again.”
I’ll go out on a limb and say that former President George W. Bush probably wasn’t thinking about Geographic Targeting Orders when he famously “improved upon” that old proverb. But after FinCEN’s recent adventures in administrative law, it feels surprisingly apropos.
On July 13th, the Ninth Circuit affirmed a preliminary injunction against FinCEN’s Southwest Border Geographic Targeting Order (GTO), which had required money services businesses in designated border communities to file Currency Transaction Reports on cash transactions exceeding just $200. In Novedades y Servicios, Inc. v. FinCEN, the court concluded that the challengers were likely to succeed on several Administrative Procedure Act (APA) claims. Most importantly, the court found the GTO was likely not really an “order” at all, but a de facto rule that should have gone through notice-and-comment rulemaking.
An important distinction, to be sure, because the GTO didn’t target particular institutions based on individualized facts. Instead, it prospectively imposed legal obligations on every covered money services business across 30 ZIP codes, rested on generalized concerns about illicit finance, and established policy applicable to an entire class of businesses. And the Ninth Circuit was unusually blunt about the significance here: the fact that a GTO applies to a “regulatory category” rather than “specific, identified businesses” is “precisely what makes it a rule.” Neither limiting that rule geographically nor giving it a 180-day expiration date changes the analysis.
The court also found that FinCEN likely acted arbitrarily and capriciously (which any lawyer will tell you is never a phrase you want to hear in a legal opinion) by failing to consider compliance costs. And the paper trail here did little to inspire confidence, to say the least.
In fact, FinCEN’s principal evidence that it had actually considered the issue was an undated, redlined draft labeled only “March XX,” with blank clearance and approval dates. When the district court asked whether it might simply have been an “after-the-fact rationalization and quick job to justify what was done,” government counsel could offer little beyond the presumption of regularity – admitting that he had no factual explanation for the “XX” and saying, in a telling moment of candor – “I don’t know what [that] means.”
Needless to say, this isn’t exactly the most ironclad legal defense, and the Ninth Circuit refused to treat the memo as a justification for the GTO.
Fast forward to August 7th – less than four weeks after Novedades – when FinCEN renewed a separate GTO aimed at government-benefits fraud in Minnesota. The order requires banks and money transmitters in Hennepin and Ramsey Counties to report additional information concerning covered international funds transfers of $3,000 or more. FinCEN had previously narrowed portions of the original order through exemptive relief, but Treasury says the renewed GTO’s reporting requirements are otherwise unchanged.
Now, to be clear, no two GTOs are built alike, and these two are hardly identical. Minnesota’s $3,000 threshold is considerably more restrained than the Border GTO’s original $200 threshold, and the Minnesota order is more specifically focused on outbound international transfers. Those distinctions just may just give FinCEN a better argument that the reporting burden is proportionate to the investigative objective (presuming, of course, FinCEN remembered to document its reasoning before issuing the order, rather than asking its lawyers to reconstruct it afterward).
It should be noted that Congress does expressly authorize GTOs for up to 180 days and permits their renewal under 31 U.S.C. § 5326(d), but that doesn’t do much to resolve the rule-versus-order problem, because as the Ninth Circuit emphasized, rules can be geographically and temporally limited too.
And even with those distinctions, there’s still the whole APA problem of it all.
Like the Border GTO, the Minnesota order doesn’t adjudicate misconduct by identified institutions. It prospectively regulates an entire category of banks and money transmitters based primarily on where they happen to operate. The geography is particularized, but the regulated parties aren’t – and according to the Ninth Circuit, it is the latter that matters. The court specifically rejected the idea that a geographically limited directive necessarily looks like an “order” where it instead imposes prospective obligations on a regulatory category of businesses.
So, said differently – take that analysis, apply it here – and things start to look awfully familiar. Congress gave Treasury separate mechanisms for imposing reporting requirements – regulations on the one hand, and § 5326 orders on the other. The Ninth Circuit specifically rejected an interpretation under which FinCEN could sidestep that distinction and impose new reporting obligations on “any category of regulated businesses in any part of the country” simply by labeling the action an “order.” Otherwise, the agency could effectively obtain rulemaking power through § 5326 without undertaking rulemaking at all.
With Minnesota, what began as a 180-day GTO in February has now simply been renewed for another 180 days, with Treasury expressly stating that (aside from previously granted exemptions) the reporting requirements remain unchanged. A supposedly temporary investigative order is therefore beginning to look markedly more like an ongoing regulatory reporting program, maintained six months at a time.
(As an aside – and as an incidental bit of irony – I’d be remiss if I didn’t mention that the Federal Register’s Official Site categorizes both the original Minnesota GTO and its renewal as a “Rule,” even though FinCEN identifies each action as an “Order.” Of course, that classification isn’t legally controlling, but FederalRegister.gov routinely categorizes actual agency orders as “Notices,” making this particular choice just a little too on-the-nose to go without comment.)
These factors, by themselves, don’t mean the Minnesota GTO is automatically unlawful. Novedades binds the Ninth Circuit, while Minnesota sits in the Eighth. Nor did the court hold that FinCEN cannot use § 5326 to impose meaningful reporting obligations.
But you simply can’t ignore the timing – FinCEN has now been explicitly warned by a federal appellate court that courts will examine the substance of agency action, not its “self-serving label,” and that § 5326 cannot become a shortcut for imposing new reporting obligations on whole categories of businesses without rulemaking. Yet only weeks later, the agency renewed another broad, prospective reporting framework governing an entire category of financial institutions, again by calling it an “order.”
So, who exactly is at risk of getting fooled again? After Novedades, a Minnesota court may eventually be asked to provide the answer.
Brett Goodnack, JD, CAMS
Compliance Advisor
OCC, FDIC Issue New CRA Notice of Proposed Rulemaking
The writing has been on the wall for the 2023 CRA rule for some time now – and at this point, it isn’t so much handwriting as it is graffiti. After all, anyone keeping tabs on the rule’s trajectory likely knows that it was enjoined before it ever took effect – and last July, the OCC, Federal Reserve and FDIC jointly proposed rescinding it and returning to the prior framework.
For all their efforts, the agencies haven’t explicitly repudiated the rationale for modernizing the CRA. Quite the opposite, in fact: in their 2025 rescission proposal, they expressly say that their “reasons for modernizing the CRA framework remain valid,” and in the new proposal, the agencies continue to acknowledge both significant changes in modern banking and “specific and known challenges” with the existing rules.
And yet, rather than merely “returning to the familiar 1995-era framework” (a phrase often repeated by the agencies throughout this process), the new NPRM would seemingly do more than restore the status quo, materially loosening several of its existing requirements – most notably through increased asset thresholds.
Currently, banks above approximately $1.649 billion generally fall into the large-bank CRA regime, which means separate lending, investment, and service tests, along with mandatory collection and reporting of certain small-business, small-farm, community-development and assessment-area data. The proposal would raise that threshold to $10 billion. Banks between $1 billion and $10 billion would instead become “intermediate banks,” subject to the substantially more flexible framework currently used for intermediate small banks. The agencies themselves concede that the large-bank tests are “more comprehensive in scope” and that the change is specifically intended to “relieve regulatory burden.”
The numbers underscore just how consequential that shift would be. Using 2024 and 2025 Call Report data, the agencies estimate that only 86 banks – or 2.4% of the approximately 3,577 OCC- and FDIC-supervised banks subject to the CRA – would remain large banks. Approximately 97.6% would fall below the $10 billion threshold.
And though we won’t delve into every substantive change of the NPRM (a full Compliance Hub summary of the proposal is currently underway and will be available shortly), it’s important to note that the relief here doesn’t stop at reporting. Intermediate banks would no longer receive separately rated investment and service tests. Arguably more striking, the proposal would allow an intermediate bank to receive an overall “satisfactory” rating based on satisfactory lending performance even where its community-development performance is weaker. The NPRM states that the change is intended to permit stronger lending performance to “compensate for weaker performance on the CD test.”
Further, banks between $412 million and $1 billion would benefit from a separate easing of the framework: because the small-bank threshold would rise to $1 billion, those institutions would lose the separate community-development test entirely.
Compliance Hub’s full summary of the NPRM will be available very soon, and for any other CRA questions, feel free to contact us on the Hotline.
Brett Goodnack, JD, CAMS
Compliance Advisor
FinCEN’s Student-Aid Fraud Alert Deserves a Closer Look
If you’re anything like me, dear reader, you may have some personal familiarity with federal student aid – and the financial reality that it leaves behind. Don’t get me wrong – years later, I’m still grateful for the education, and for the jumble of letters it tacked onto the end of my name. But while my diploma might be gathering dust, the debt it helped finance remains considerably more active. Student debt can shape a person’s finances for years, which makes fraud involving the aid system a genuinely serious concern -and makes imprecision surrounding that fraud just as serious.
Which brings us to FinCEN’s latest alert on federal student-aid fraud, which identifies a real and increasingly sophisticated threat, but also appears to blur important technical distinctions. To that end, the alert outlines “ghost student” schemes in which fraudsters use stolen or synthetic identities to enroll in colleges, obtain federal aid, and pocket the remaining funds after tuition is paid. It also identifies complicit “straw students,” corrupt school employees, money mules, shell companies, and accounts receiving refunds intended for unrelated individuals. Practically speaking (and as is often the case with FinCEN Alerts) the most useful portions of the document for banks are likely the transactional red flags: multiple student-aid refunds entering the same account, newly opened accounts funded only by refunds, rapid transfers through peer-to-peer services, and funds quickly converted into digital assets or sent abroad.
So, the alert certainly isn’t “crying wolf.” The alert may, however, be describing that wolf in unusually…let’s call them, ”cinematic”…terms. FinCEN repeats the Department of Education’s claim that it prevented more than $1 billion in student-aid fraud during 2025, but neither the alert nor the cited Department release provides enough methodology to independently evaluate that figure. The Department points to nearly 150,000 “suspect identities,” yet doesn’t venture to explain how many were confirmed as fraudulent, how much aid was associated with each application, how false positives or duplicate submissions were treated, or how much of the total represented money that otherwise would actually have been paid.
The alert also says students must remain enrolled for 60 percent of an enrollment period “to receive a full refund.” In reality, the 60-percent threshold generally applies when a student withdraws and determines how much Title IV aid is treated as earned under the Return of Title IV Funds calculation. A credit-balance refund may be paid earlier; remaining enrolled beyond the threshold generally affects whether that aid must later be returned.
More broadly, the document offers surprisingly little aggregate data for an alert described as being derived in part from Bank Secrecy Act reporting. There aren’t any SAR totals, aggregate dollar trends, geographic concentrations, review period, or meaningful quantitative breakdown among domestic, foreign, ghost-student, straw-student, and insider-assisted schemes. Yet it invokes foreign actors, digital assets, senior foreign political figures, correspondent banking, terrorism reporting, and other subjects that aren’t really tied to any student-aid-specific data presented in the alert.
And, so, I’m almost forced to reach a conclusion that I’ve arrived at a few too many times in the recent past—it presents the threat with considerably more sweep than the disclosed data can quantitatively support. Why gild the lily? Student-aid fraud already carries enough weight. And frankly, it doesn’t take a jumble of letters after your name to recognize the difference between evidence and embellishment.
Brett Goodnack, JD, CAMS
Compliance Advisor