Ask a room of risk officers whether their financial institution (FI) follows an AI risk management framework, and you'll get three answers: yes, no, or "which one?" The third answer is the most honest one, and it points straight at the problem. FIs have six AI risk management documents to work with:
- The NIST AI RMF
- The FS AI RMF
- ISO/IEC 42001
- SR 26-2/OCC 2026-13
- COSO's generative AI guidance
- CSBS AI Supervisory Framework
None of them is a binding rule.
The AI governance space went from empty to crowded in about three years. Every new document comes with a big acronym, a triple-digit-page count, and dense language. FIs can't tell which document is a rule, which is a suggestion, which one an examiner cares about, and which two are saying the same thing in different dialects.
This guide covers what each of the six documents is, which ones apply to your FI, where to start and how to choose the appropriate one, and how enterprise risk management software can support your AI governance program.
Related Webinar: AI Demystified: What Every Financial Institution Leader Needs to Know Right Now
What Is an AI Risk Management Framework?
An AI risk management framework sets out the outcomes an FI should try to achieve and leaves implementation to the FI. That's what separates it from the other two kinds of documents in this space: standards and guidance. An AI standard specifies requirements you can be audited against, while AI guidance tells you what your regulator or a standard-setting body expects to see.
All six AI risk management docs are casually referred to as AI frameworks, but only two of them are. That label decides what you most need to know about each one.
Related: Laws vs. Regulations vs. Rules vs. Guidance: What Are the Differences?
The 6 AI Risk Management Documents Financial Institutions Should Know
Of the six AI documents most often used by FIs (NIST AI RMF, FS AI RMF, ISO/IEC 42001, SR 26-2/OCC 2026-13, CSBS AI Supervisory Framework, and COSO GenAI), none carries the force of law. Still, they are all helpful tools for governing your AI risk sensibly.
The table below compares all the AI risk management documents side by side, including what each one is and whether it covers generative AI (GenAI).
| Document | What It Is | Issued by | Published date | In brief |
Covers generative AI |
| NIST AI RMF: AI Risk Management Framework 1.0 (NIST AI 100-1) |
Framework | NIST | January 2023 | The voluntary U.S. baseline. Sector-agnostic and outcome-based, structured around four functions (Govern, Map, Measure, Manage) | Technology-agnostic; can include it |
| FS AI RMF: Financial Services AI Risk Management Framework | Framework | Cyber Risk Institute (CRI) & FSSCC | February 2026 | NIST's framework operationalized for financial services, adding ~230 sector-specific control objectives scaled to an FI’s AI maturity | Yes |
| ISO/IEC 42001:2023: AI management system | Standard | ISO/IEC | December 2023 | A formal, certifiable management-system standard for governing AI. It’s the governance wrapper around your AI models, not the models themselves | Technology-agnostic; can include it |
| SR 26-2 / OCC 2026-13: Revised Interagency Guidance on Model Risk Management | Supervisory guidance | Federal Reserve, OCC, FDIC | April 2026 | The regulators' revised model-risk discipline for traditional models, covering development, validation, and effective challenge across the lifecycle | Excluded |
| COSO GenAI: Achieving Effective Internal Control Over Generative AI | Thought leadership guidance | COSO | February 2026 | Established internal-control discipline (COSO's ICIF) applied specifically to generative-AI risks, organized by what the technology does | Written for it |
| CSBS AI Supervisory Framework: Examination Framework for State Examiners | Supervisory guidance | CSBS | September 2026 | Built directly on NIST AI RMF and FS AI RMF; an examiner’s playbook for state-chartered banks and state-licensed non-banks |
Yes |
NIST AI RMF 1.0: The Voluntary U.S. Baseline
The AI Risk Management Framework 1.0 (NIST AI 100-1), published January 26, 2023, is the reference point for almost everything that followed. It answers the first question any FI has to settle: what outcomes should we be aiming for in our AI oversight and governance? It's voluntary, sector-agnostic, and non-certifiable, so you self-attest alignment. There's no NIST audit or certificate. Its architecture is centered on four core functions: Govern, Map, Measure, and Manage. Govern runs through all the others rather than sitting as a first step you finish and forget.
Beneath those functions sit 19 categories and 72 subcategories. NIST also defines seven trustworthiness characteristics: valid and reliable, safe, secure and resilient, explainable and interpretable, privacy-enhanced, fair with harmful bias managed, and accountable and transparent. These are outcomes to aim for, not a checklist to work through in order. Think of it as the shared vocabulary the rest of the field borrows.
Related: What Bankers Need to Know About NIST 2.0
FS AI RMF: Translated for Financial Services
If NIST wrote the grammar standards, the Financial Services AI Risk Management Framework wrote the industry dictionary. The Cyber Risk Institute and the Financial Services Sector Coordinating Council released it on February 19, 2026, under a Treasury-led oversight group. It operationalizes NIST for financial services, keeping the four-function skeleton and adding a layer underneath: roughly 230 control objectives defining the sector-relevant outcomes an FI can implement and evidence.
It's voluntary, industry-led, and non-certifiable, but it does two things NIST doesn't. First, an adoption-stage questionnaire (Initial, Minimal, Evolving, Embedded) points you to the controls that fit your AI maturity, so you're not working through all 230. Second, it explicitly covers GenAI and agentic AI.
Related: Emerging Risks in Banking: Q3 2026 Update
ISO/IEC 42001: The Only Certifiable AI Standard
Published December 18, 2023, ISO/IEC 42001 is the world's first certifiable management-system standard for AI. Of all the documents here, it's the only one an accredited body can certify, which is its main differentiator.
It defines the requirements for an AI Management System (AIMS), the governance wrapper you run around your AI rather than the models themselves. The structure follows ISO's standard management-system format, the same skeleton behind ISO 27001, so it slots in next to systems you may already run. The auditable requirements live in Clauses 4 through 10. Clause 6 holds the risk engine, which calls for an AI risk assessment, risk treatment against a reference set of 38 Annex A controls, and an AI impact assessment that weighs consequences to individuals and society alongside consequences to your FI.
SR 26-2 / OCC 2026-13: Model Risk Management Guidance
Issued April 17, 2026, by the Federal Reserve, OCC, and FDIC, the Revised Interagency Guidance on Model Risk Management is the one most likely to feel official, since it comes from regulators. It's supervisory guidance rather than a rule. It's also principle-based and proportionate, so you scale your model-risk rigor to each model's materiality and to your institution's size and complexity. Primary relevance is flagged for Fed-regulated banking organizations with over $30 billion in total assets. It supersedes SR 11-7 and SR 21-8, consolidating over a decade of supervisory experience into one document.
Two things about it matter most. It governs the full lifecycle of traditional models, including conventional machine learning, through development, validation, and effective challenge. It also carves generative and agentic AI out of scope, leaving those to a planned interagency request for information. The carve-out is deliberate, and COSO GenAI fills that seam.
COSO GenAI Guidance: Internal Control for Generative AI
Released February 23, 2026, COSO's Achieving Effective Internal Control Over Generative AI is the odd one out, and on purpose. It's an internal-control and audit lens rather than a risk framework in the NIST sense. It takes COSO's Internal Control-Integrated Framework (ICIF), the same five components and 17 principles underpinning most U.S. internal control over financial reporting, and applies them to GenAI risks such as hallucination, drift, and prompt injection.
It has a capability-first taxonomy that organizes GenAI by what the technology does, from ingestion and transformation through orchestration, judgment, and human-AI interaction, and attaches control expectations to each. A six-step roadmap wraps the whole thing: govern, inventory, assess, design, implement, monitor. If SR 26-2 is the discipline your traditional models run through, COSO GenAI is the discipline your generative AI runs through, in the control language your ICFR program already speaks.
CSBS AI Supervisory Framework: The New State Exam Playbook
Released September 16, 2026, the CSBS AI Supervisory Framework is the newest document in this space, and the first built specifically for state examiners. It isn’t a new vocabulary or a new set of controls: it’s built directly on NIST AI RMF and FS AI RMF, so it tests the same programs covered above rather than introducing new requirements.
It applies to state-chartered banks and state-licensed nonbanks; each state decides whether to adopt it, so expect variation, and credit unions are not in scope. Its most important feature is that it explicitly covers GenAI and agentic AI, precisely the space SR 26-2 carves out at the federal level, closing that gap at the state level until federal regulators weigh in through their planned interagency request for information (RFI).
Are AI Risk Management Frameworks Legally Binding?
None of the six is a binding rule, but non-binding covers a lot of ground. NIST AI RMF is a voluntary baseline. FS AI RMF is an industry framework. ISO 42001 is a certifiable standard. SR 26-2 and OCC 2026-13 are supervisory guidance from your regulators. COSO GenAI is internal control guidance. CSBS’s AI Supervisory Framework is a discretionary state examination tool. Voluntary doesn't mean unimportant, and the differences between these categories determine how you demonstrate that you're doing the work, and to whom.
Which AI Frameworks Cover Generative AI?
Generative AI is the fault line that organizes the whole field. SR 26-2 and OCC 2026-13 exclude generative and agentic AI. COSO GenAI was written specifically for it. FS AI RMF covers it. ISO 42001 and NIST are technology-agnostic enough to stretch across it. CSBS AI Supervisory Framework explicitly puts generative and agentic AI back on the table at exam time.
What Each AI Risk Management Document is For
Most FIs don't have an AI framework problem. They have a categorization problem, and that one's solvable.
The six AI risk management documents are six instruments answering six different questions: What outcomes should we aim for? How do we operationalize that in financial services? How do we get certified? How do we handle traditional model risk? How do we control generative AI specifically? How will a state examiner test it?
AI governance starts with sorting out which question each document answers. It's also the part you can do today, without waiting for the next round of guidance. Answer that, and you have a program you can defend rather than a stack of documents nobody can explain.
Where Should a Financial Institution Start?
When choosing an AI risk management document, start with the foundation. NIST AI RMF is the shared vocabulary, and FS AI RMF operationalizes it for financial services with concrete control objectives scaled to your AI maturity. Between the two, that’s your foundation — not one choice among several, but the program everything else attaches to.
From there, what you add depends on the question you’re trying to answer:
If your federal examiner is the audience: SR 26-2 and OCC 2026-13 are the documents your Federal Reserve, OCC, or FDIC examiner will have read. They govern traditional model risk for supervised banking organizations.
If your state examiner is the audience: The CSBS AI Supervisory Framework is built directly on NIST AI RMF and FS AI RMF, so a state examiner is testing the program you’ve already built, not a new one.
If you need to demonstrate governance externally: ISO 42001 is the only one that produces a certificate an accredited body can issue. It's the document to reach for when you need to show something independently verified.
If generative AI is the immediate concern: COSO GenAI offers a practical, COSO-aligned approach to managing the risks and opportunities introduced by rapidly advancing generative AI technologies, in the internal-control language your audit function already speaks.
How much of each you need — and how to size that combination to your institution rather than treat all as mandatory — is its own question.
How Risk and Compliance Management Software Supports AI Governance
Every one of these documents ends in the same place: documented assessments, controls someone can inspect, and evidence you can give an examiner. That's where risk and compliance management software can support your FI.
Nrisk, Ncontracts’ risk management software for FIs has more than 100 pre-built, configurable risk assessment templates, including an AI risk assessment template, and hundreds of expert-built controls, so you're not starting from a blank page. Compliance management software carries AI and model risk management policies. Audit management software adds an AI audit template for the internal control and audit lens COSO GenAI is built around. Vendor risk management software includes an AI questionnaire for the exposure that arrives through your vendors rather than your own builds.
Once you know which document you're working from, the next question is how to review your AI systems and tools. Our free AI Auditing Checklist, built on the Three-Domain Approach, walks through governance, management, and internal audit oversight.
