In this article
If your examiner asks how your institution governs the chatbot answering member questions or the assistant drafting credit memos, what document do you point to? For most community banks and credit unions, the honest answer is none, because the new federal model risk guidance deliberately left generative and agentic AI out. The Treasury Financial Services AI Risk Management Framework is the closest thing to an answer, and it is more usable than most compliance teams assume.
Why Generative AI Has No Federal Bank Rulebook Yet
On April 17, 2026, the OCC, the Federal Reserve, and the FDIC replaced the 2011 model risk guidance with revised interagency guidance, and the OCC's Bulletin 2026-13 is blunt about scope: "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance." The same bulletin says the agencies plan to issue a request for information on banks' use of AI, including generative and agentic AI, and that the guidance is expected to be most relevant to banking organizations with more than $30 billion in assets.
We covered what that exclusion means for exam preparation in our earlier look at how sector regulators kept moving while state AI laws stalled. This piece picks up where that one stopped: once you accept that the burden sits with your own governance program, which framework should that program actually be built on?
Credit unions face the same question from a different angle. The NCUA states that it has issued no AI-specific rules and supervises AI through its existing safety-and-soundness, compliance, internal control, and third-party due diligence expectations. Notably, the same NCUA page points credit unions to the resources Treasury developed with industry, which tells you where a credit union examiner expects you to look.
What Is The Treasury Financial Services AI Risk Management Framework?
The Treasury Financial Services AI Risk Management Framework, or FS AI RMF, is a voluntary, financial-sector adaptation of the NIST AI Risk Management Framework that gives institutions 230 control objectives scaled to how deeply they use AI. Treasury released it on February 19, 2026, alongside a shared AI Lexicon, and it does not carve generative or agentic AI out of scope. For a community bank or credit union, it is the most specific federal reference available for governing those tools today.
According to Treasury's announcement, the framework was developed through the Financial and Banking Information Infrastructure Committee and the Financial Services Sector Coordinating Council's AI Executive Oversight Group, and Treasury describes it as scalable for institutions of varying size and complexity. The Cyber Risk Institute, which maintains the framework, says it was built with more than 100 financial institutions and is intended for community banks, credit unions, national banks, insurers, investment firms, and their third-party providers.
Here is the belief worth challenging. Many compliance officers treat a 230-objective framework as something only a large bank can afford to adopt, and they file it away until a regulator makes it mandatory. That reads the design backwards. The framework is organized by adoption stage precisely so a smaller institution with modest AI use can work from a much shorter list, and waiting for the binding version means building your program under exam pressure instead of on your own schedule.
How Does The Framework Scale To A Smaller Institution?
It scales through an adoption-stage questionnaire that determines which subset of control objectives applies before you touch the full matrix. The FS AI RMF Adoption Stage Questionnaire places an institution in one of four stages, assessed across six dimensions: business impact, governance, deployment model, third-party AI use, organizational goals, and data sensitivity and criticality.
| Adoption stage | How the questionnaire describes it | What it typically looks like at a community institution |
|---|---|---|
| Initial | Limited use; AI not embedded in critical functions or business decisions | Staff experimenting with a general assistant under an acceptable use policy |
| Minimal | Low-risk implementations for non-critical tasks | Meeting summaries, internal drafting, help desk triage |
| Evolving | AI driving outcomes through external-facing solutions and sensitive data, but not critical decision-making | A member-facing chatbot, AI features in the core or digital banking platform |
| Embedded | AI driving outcomes in critical functions through autonomous decision-making | AI influencing credit, fraud, or account decisions with limited human touch |
Each stage maps to its own workbook within the Risk and Control Matrix, so the stage you land on determines which control objectives you work through first. The other two components matter as much as the matrix. The Guidebook explains how to operationalize the controls, and the Control Objective Reference Guide gives examples of controls and the evidence that supports them, which is the part examiners and auditors actually ask to see.
The stage assessment is also where most institutions discover they are further along than leadership thinks. A vendor quietly switching on a generative feature in the online banking platform can move an institution from Minimal to Evolving without anyone filing a project request. That is why the third-party AI use dimension deserves the most honest answer on the questionnaire.
The Five Other Deliverables Most Teams Skip
The framework was one of six coordinated deliverables, and the FSSCC's AIEOG deliverables page lists the rest: the AI Lexicon, identity and authentication guidance on AI-powered attacks, an explainability paper focused especially on generative AI in finance, a Data Nutrition Labeling approach for evaluating data quality, and an AI-enhanced fraud resource covering incident response and deepfakes. A program built on the framework alone misses the companion pieces that answer the questions examiners are most likely to ask.
The fraud and identity work is the most urgent of these for smaller institutions. Help Net Security's coverage of the industry identity paper reports Deloitte's Center for Financial Services projection that U.S. AI-enabled fraud losses could reach 40 billion dollars by 2027, up from 12.3 billion dollars in 2023. Community institutions see that pressure first at account opening and in social engineering against staff, the same tactics we break down in how cybercriminals use AI to launch smarter attacks.
Explainability connects to a duty you already have. Under Regulation B's adverse action notice requirements, a creditor that takes adverse action must provide the specific principal reasons or disclose the applicant's right to request them. If an AI tool contributes to a credit decision, someone at your institution has to be able to state those reasons accurately, and the explainability paper is written for exactly the non-technical business owners who carry that responsibility.
Where Should A Bank Or Credit Union Start?
Start with an inventory and the questionnaire, then work only the control objectives for your stage. Institutions that try to read all 230 objectives first usually stall, while those that pin down their stage and their actual AI footprint can produce a defensible first pass in one review cycle.
A practical first sequence looks like this:
- Build an inventory of every AI capability in use, including features vendors enabled inside platforms you already license, and record which ones touch member or customer data.
- Run the adoption stage questionnaire with risk, compliance, IT, and the business line owners in the same room, so the third-party answers are not guesses.
- Work the matrix for your stage and assign a named owner to each applicable control objective, with the evidence the Reference Guide suggests.
- Extend vendor due diligence to ask how each provider's AI features are tested, monitored, and changed, and what audit rights you hold.
- Set a quarterly review, and treat any vendor AI release note as a trigger to revisit your stage.
The vendor step is where most of the real exposure sits. Most generative AI enters a community institution through a core processor, a digital banking provider, or a fraud tool rather than an internal build, and the institution still owns the risk. Our guidance on the AI vendor contract clause that decides whether you can prove compliance covers the contract language, and the broader discipline is the same one described in protecting your business from vendor vulnerabilities.
It also helps to set a bright line on what AI may do on its own before any control work begins. Our position, laid out in the bright line for AI in finance, is that AI can assist and flag but should not move money or finalize a consequential decision without a person. That single decision keeps many institutions out of the Embedded stage entirely, which shrinks the control set they need to build.
How Securafy Helps Institutions Put The Framework To Work
From a compliance standpoint, the difficult part of the FS AI RMF is not understanding the control objectives. It is producing evidence that holds up when an examiner asks for it. Securafy starts with discovery across the environment, including the AI features vendors have switched on inside Microsoft 365, the core, and the digital banking stack, then runs the adoption stage assessment with your team and maps each applicable control objective to an owner and an evidence source.
Because we already manage security and IT operations for financial institutions, much of that evidence comes from systems we can already see: access logs, data loss prevention policies, vendor reviews, and change records. The result is a governance file that ties your AI use to the Treasury framework and your existing vendor management program, so it reads as part of how the institution operates, the approach we argue for in why AI adoption is a governance problem, not a technology one.
Where To Go From Here
The interagency rulebook for generative and agentic AI is still on the way, but your examiner's questions are not waiting for it. Run the adoption stage questionnaire this quarter, and you will know which part of the Treasury framework your institution actually needs to build.
If your team is moving faster with AI than your guardrails are, start with structured training rather than another tool. AI University gives your people role-based AI training with security built into the material, not bolted on afterward.
If you would rather talk through your specific environment first, book a strategy call with Securafy and we will walk your current AI usage, exposure, and the fastest path to safe adoption.
Need help with AI governance? Adapt the AI Acceptable Use Policy Template with your legal and IT teams.
Get the template
Ric Hall is the Chief Revenue Officer at Securafy, with decades of experience in enterprise infrastructure, cloud technology, sales leadership, and business strategy.
He writes for leaders trying to make sense of big technology decisions without getting trapped in vague promises or polished sales language. His articles cover provider selection, IT budgeting, co-managed services, cybersecurity investments, modernization, and the questions businesses should ask before signing a contract.
Ric’s strength is connecting technical decisions to business outcomes, helping leaders understand not just what they are buying, but why it matters and whether it will still make sense 3 years from now.
Writes about: IT budgeting, provider evaluation, cybersecurity ROI, co-managed IT, cloud modernization, vendor selection, technology strategy
Join the conversation
Have a question or a different take on this? Add it below.