AI governance framework guide
AI governance framework: components and models
What an AI governance framework contains, how the published frameworks and models compare, and how a regulated mid-market company can adapt one to the rules it already answers to.
An AI governance framework is the set of roles, rules and records an organization uses to decide which AI systems it runs, how each one is tested and monitored, and who answers when one goes wrong. Published models such as NIST AI RMF 1.0 and ISO/IEC 42001 describe what to cover; your framework applies them to your systems. An artificial intelligence governance framework is the same thing under its full name.
what is an ai governance framework
What is an AI governance framework?
A policy says what you believe about AI. A framework says who decides, on what evidence, and where the decision is written down. If your organization has an AI policy but cannot list every AI system it runs, name the owner of each, and show when each was last checked, it has a policy, not a framework.
A working AI governance framework answers five questions for every AI system: what it is and who owns it, how risky it is, what testing it passed before it went live, how it is watched now, and who can pause or retire it. Everything else on this page is a way of answering those five questions so the answers survive staff turnover and a board review.
The words around this topic get used loosely. In this guide, a framework is your organization's own operating model. A published model or standard, such as the NIST AI Risk Management Framework or ISO/IEC 42001, is a reference you build from. Broader questions about what AI governance is and why it matters are covered in our AI governance consulting guide, and the line between governance and compliance is drawn in AI governance vs AI compliance.
context
What does an artificial intelligence governance framework cover?
An artificial intelligence governance framework covers four things: accountability, risk, data and evidence. Accountability means a list of the AI systems in use, an owner for each, and a named person who approves new uses. Risk means a rating for each system and a plan for how it is tested and watched. Data means which data each system may use and who can reach it; our AI data governance guide covers that. Evidence means the records of approvals, test results and exceptions that a board or auditor can review. The nine components below break these down.
context
What is AI risk management?
AI risk management is the practice of finding what could go wrong with each AI system you run, deciding how serious it is, and putting an owner and a control against it. In an AI governance framework it shows up as a risk tier for every system, testing before go-live, monitoring afterwards, and a named person who can pause the system. Published models such as the NIST AI RMF 1.0 describe what to cover; your framework applies them to your own systems. We recommend starting with an inventory, because you cannot rate the risk of a system nobody has listed. We also recommend sizing the controls to the risk, with each tier based on the decision the system influences and who it affects, attaching heavier testing and review only to the higher tiers.
context
What is the difference between governance and risk?
The difference between governance and risk is that governance decides who is accountable and which rules apply, while risk work finds and weighs what could go wrong. For AI, governance is the owners, rules and approvals that say how a system may be used. Risk work asks what harm it could do, how likely that is and how you would notice. Neither works alone: rules nobody has tested are only a policy, and risk findings with no owner go unfixed. NIST describes its AI Risk Management Framework as voluntary and groups the work into four functions: Govern, Map, Measure and Manage. In our reading, Govern is the governance side and the other three are risk work; NIST's pages do not put it that way. NIST's framework page says AI RMF 1.0 is being revised, and its Playbook page says the Playbook will be updated after that revision.
framework components
The components of an AI governance framework
A working framework has nine components. Each one produces a record, not just intent, and each has a named owner rather than living only in a policy document. Items 1-5 are the building blocks already on this page; items 6-9 complete the set.
1. Model and use-case inventory
Every AI system in production or pilot, cataloged with its owner, business purpose, data sources, and a documented risk tier based on the decision it influences and who it affects. An inventory that only covers models built in-house misses the vendor tools and embedded copilots that carry the same risk. The data those systems use needs owners too, and our page on data governance solutions shows what the tooling does.
2. Intake and approval gate
A single point where a new AI use case gets registered before it touches real data or a real decision, so the governance committee is never finding out about a system after it is already live. The gate assigns a risk tier and routes higher-tier systems to deeper review.
3. Validation and testing standards
Tier-appropriate testing before deployment: for models that can be backtested, statistical performance and bias testing against a holdout set; for generative and agentic systems, structured output sampling against a rubric, red-teaming for guardrail failures, and adversarial testing of the tools an agent is allowed to call.
4. Ongoing monitoring and drift controls
Ownership of a system does not end at go-live. Monitoring tracks output quality, data drift, and guardrail failures on a cadence tied to risk tier, with a defined path to pause or roll back a system that drifts outside its validated behavior. Our short guide to model drift detection explains how the baseline comparison behind those checks works. Running that cadence in production, with deployment, drift monitoring and runbooks handled for you, is what our MLOps governance as a service covers.
5. Documentation and audit trail
Every decision the framework makes, an intake approval, a validation result, a monitoring exception, a model retirement, is written down in a form a regulator, auditor, or board member can review without asking the original team to reconstruct it from memory. Agents need the same record for every tool they call, which we cover in why AI agent deployments fail governance audits.
6. Roles and decision rights
Write down who can approve a new AI system, who can pause one, and who signs off on exceptions. An AI Center of Excellence, or CoE, is the operating team that runs the framework day to day; it is where the technical judgment lives. A governance committee holds the authority to approve, pause or retire a system. In a mid-market company the CoE can be two or three people from existing staff. Our AI Center of Excellence governance guide covers committee charters.
7. Policies and acceptable use
The written rules staff follow day to day: which AI tools are approved, what data may go into them, and what needs a human check before it leaves the building. The policy is one input to the framework, not the framework itself. Our guide to AI policies lists the policies a company usually needs. Unapproved tools already in use belong in the inventory too; see our guide to managing shadow AI risk.
8. Third-party and vendor AI
Vendor models, AI features inside software you already license, and outsourced builds go through the same intake and tiering as anything built in-house. Record what the vendor has told you in writing, what you tested yourself, and what you could not test. Our AI vendor evaluation checklist lists the questions to ask.
9. Training and review of the framework itself
People who approve, build or use AI systems are trained on the rules that apply to them, and the framework has an owner who reviews it on a set schedule and whenever a new kind of system or new guidance arrives.
frameworks compared
AI governance frameworks and models compared
Several published AI governance frameworks and standards are available to build from. They differ in who wrote them, whether they are voluntary, and how much structure they give you. Each row below describes the document from its own publisher's text, linked in the sources at the foot of the page.
| Framework or model | Published by | What it says it is | How it is organized | Useful to you as |
|---|---|---|---|---|
| NIST AI Risk Management Framework (AI RMF 1.0) | US National Institute of Standards and Technology, released January 2023 | A framework 'intended for voluntary use', described by NIST as non-sector-specific and use-case agnostic | Four functions: Govern, Map, Measure and Manage. NIST states that Govern applies across all stages, while the other three apply to specific systems and lifecycle stages. A companion Playbook and a Generative AI Profile are published alongside it | A detailed backbone for your control list and risk vocabulary |
| ISO/IEC 42001:2023 | ISO and IEC, 2023 | An international standard for an AI management system | A standard for an AI management system, issued by ISO and IEC | A structure for running AI governance as a management system, alongside other management systems you may already run |
| EU AI Act | European Parliament and Council, Regulation (EU) 2024/1689 of 13 June 2024, since amended by Regulation (EU) 2026/1744 | A regulation 'laying down harmonised rules on artificial intelligence' | Its subject matter includes prohibitions of certain AI practices, specific requirements for high-risk AI systems and obligations for their operators, and transparency rules for certain AI systems | A reference for risk-based classification, and a matter for your counsel if you place AI systems on the EU market or use them there |
| OECD AI Principles | OECD, adopted May 2019 and updated May 2024 | Principles that promote use of AI that is innovative and trustworthy and that respects human rights and democratic values | Five values-based principles (inclusive growth, sustainable development and well-being; human rights and democratic values, including fairness and privacy; transparency and explainability; robustness, security and safety; accountability) plus recommendations for policymakers | A source for the values statement at the top of your framework, and for a definition of 'AI system' |
| Model AI Governance Framework (Singapore) | Singapore's Personal Data Protection Commission (PDPC); first edition January 2019, second edition January 2020 | Guidance for organisations that turns ethical principles into practices they can adopt when deploying AI; sector- and technology-agnostic | Two guiding principles (decisions made by AI should be explainable, transparent and fair; AI systems should be human-centric) and four areas: internal governance structures and measures; the level of human involvement in AI-augmented decision-making; operations management; stakeholder interaction and communication | A checklist of governance areas, useful for structuring roles and human oversight |
Standards, frameworks and models
ISO/IEC 42001 is a standard for an AI management system. NIST AI RMF calls itself a framework intended for voluntary use. The EU AI Act is a regulation, the OECD text is a set of principles, and Singapore's document is a model framework. An AI governance model, as this guide uses the phrase, is the shape of your own operating model: who decides, at what level, with what evidence. Your framework can borrow structure from one published document and vocabulary from another, as long as each control traces to a reason.
Sector documents
In regulated industries, sector guidance and rules may also apply to you; your counsel decides which. We do not certify compliance or say what a regulator will accept. The table describes what each named document says about itself.
| Sector | Document | What the document says |
|---|---|---|
| Banking | SR 26-2, Revised Guidance on Model Risk Management (April 17, 2026) | Issued by the Federal Reserve, OCC and FDIC. It states that it 'supersedes and replaces' SR 11-7 and SR 21-8, and emphasizes a risk-based approach tailored to a banking organization's model risk profile and the size and complexity of its operations. Its cover letter says it is expected to be most relevant to banking organizations with over 30 billion dollars in total assets regulated by the Federal Reserve, and the guidance says generative AI and agentic AI models are 'not within the scope of this guidance'. See our SR 26-2 guide. |
| Insurance | NAIC Model Bulletin: Use of Artificial Intelligence Systems by Insurers (adopted December 4, 2023) | A model bulletin that states can adopt; each state adopts its own version. The text says insurers 'are expected to develop, implement, and maintain a written program' (an 'AIS Program') for AI systems that make or support decisions related to regulated insurance practices, and that the program 'may adopt, incorporate, or rely upon' a framework such as the NIST AI RMF. See our NAIC AI model governance page. |
| Healthcare | HIPAA Security Rule (45 CFR Part 164, Subpart C) | Titled 'Security Standards for the Protection of Electronic Protected Health Information', with sections on administrative, physical and technical safeguards. See our guide to HIPAA AI compliance requirements. |
| Life sciences | 21 CFR Part 11 | Titled 'Electronic Records; Electronic Signatures'. Its scope section says it applies to records in electronic form that are 'created, modified, maintained, archived, retrieved, or transmitted, under any records requirements set forth in agency regulations'. |
choose and adapt
How to choose and adapt a framework
If your company has no model risk department or standing AI office, the framework runs on the people you already have. Aim for the smallest set of controls that covers your real systems. These steps are one sensible order.
1. Start from what you already answer to
List the sector documents, contracts and internal policies that already apply to you, and have counsel confirm the list. Then reuse the risk, audit and committee processes you already run rather than building a parallel structure. The NAIC bulletin (a model bulletin; each state adopts its own version) says the AIS Program 'may be independent of or part of' an insurer's existing enterprise risk management program.
2. Pick one backbone, borrow the rest
Choose one published framework for structure: NIST AI RMF for a detailed control vocabulary, or ISO/IEC 42001 if you already run ISO management systems. Borrow the OECD definition of an AI system and Singapore's governance areas where they help. Record which document each control came from.
3. Inventory before you write rules
Catalog every AI system in production or pilot, including vendor features and tools staff adopted on their own, with an owner for each. Do the inventory first, so the rules cover the systems you actually run.
4. Size controls to risk
Write a short tiering rubric based on the decision each system influences and who it affects, then attach heavier testing and review only to the higher tiers. Both SR 26-2 and the NAIC bulletin use proportionality language: SR 26-2 describes a risk-based approach, and the bulletin says the program 'should be tailored to and proportionate with' the insurer's use of AI.
5. Name owners, then run one real review
Name the CoE owners and the committee members, then put two or three real use cases through intake, testing and approval before the framework is finalized. Unclear rules show up the first time a real system goes through the process.
6. Write it down and set the review date
Publish the framework document, the tiering rubric and the intake form in one place, and put the next framework review on the calendar.
Who this is not for: if you run one or two low-risk internal tools and no customer-facing or regulated decisions touch AI, a written acceptable use policy and an inventory may be enough for now. If you operate across many countries with a dedicated model risk function, you likely need a larger program than this guide describes. For the general parts of any compliance program, see our plain guide to the compliance program.
If you want a structured starting point, our AI readiness and governance assessment maps your current systems against these components.
framework outline
An AI governance framework outline you can adapt
A framework document can follow this table of contents. Some teams call the operational half of it an AI governance playbook. Adapt the headings; keep the order, because each section depends on the one before.
1. Purpose, scope and definitions
Which business units and systems the framework covers. Define 'AI system' once; the OECD definition is a reasonable starting point.
2. Principles
Five to seven commitments in plain language, traceable to a published source such as the OECD principles or the NIST trustworthiness characteristics.
3. Roles and decision rights
CoE, committee, system owners, risk and compliance, security, legal. Who approves, who can pause, who is consulted.
4. Inventory and risk tiering
The fields every inventory entry carries and the rubric that assigns a tier.
5. Intake and approval
The intake form, what evidence each tier must bring, and how a decision is recorded.
6. Testing and validation by tier
Statistical testing for predictive models; output sampling, red-teaming and tool-permission tests for generative and agentic systems.
7. Monitoring, incidents and change
What is watched, how often for each tier, what triggers a pause, and how a material change sends a system back through intake.
8. Third-party AI
Vendor intake, contract and data questions, and what you test yourself.
9. Records and evidence
Where approvals, test results and exceptions are kept and for how long, set by your records policy.
10. Training and framework review
Who is trained on what, and the date and owner of the next framework review.
11. Appendix: mapping table
One row per control, with columns for the published framework it comes from (for example a NIST AI RMF function) and the sector document it supports.
The acceptable use policy is a separate, shorter document that staff read; this outline is the operating document the CoE and committee work from. Both should point at each other.
what a kriv engagement includes
What a Kriv AI Governance Engagement Includes
Assessment: a current-state review of every AI system in production or pilot, mapped against the policies and frameworks you already answer to, with gaps documented against the components above.
Framework design: the intake gate, the risk-tiering rubric, and the validation and monitoring standards, written to the enterprise's existing risk and compliance processes rather than replacing them wholesale.
CoE and committee stand-up: charter drafting, initial staffing recommendations, and running the first review cycle alongside the enterprise's own team so ownership transfers rather than staying with the consultant.
Ongoing advisory, if you want it: review of the framework as new use cases, new models, or new guidance such as SR 26-2 change what it needs to cover.
Once the framework is written, an enterprise AI governance program keeps it running through inventory, risk tiering and change review.
engagement
How engagements are scoped and priced
Every engagement is scoped and priced after a short discovery call.
Sources
Cited sources
- NIST AI Risk Management Framework (page, read 2026-10-19)
- NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0) (PDF, read 2026-10-19)
- Regulation (EU) 2024/1689 (Artificial Intelligence Act) of 13 June 2024, Official Journal, EUR-Lex (read 2026-11-01)
- OECD AI Principles overview, OECD.AI (read 2026-10-19)
- Singapore's Approach to AI Governance (Model AI Governance Framework), PDPC (read 2026-10-19)
- SR 11-7 (2011), Guidance on Model Risk Management: superseded and replaced by SR 26-2 on April 17, 2026
- SR 26-2, Revised Guidance on Model Risk Management, Federal Reserve (read 2026-10-19)
- NAIC Model Bulletin: Use of Artificial Intelligence Systems by Insurers (PDF, read 2026-10-19)
- 45 CFR Part 164 Subpart C, Security Standards for the Protection of Electronic Protected Health Information, eCFR (read 2026-11-01 via the eCFR API)
- 21 CFR Part 11, Electronic Records; Electronic Signatures, eCFR (read 2026-11-01 via the eCFR API)
- NIST, AI RMF Playbook (read 2026-11-10)
Related resources
Continue exploring
Straight answers
Frequently asked questions about the AI governance framework
What is an AI governance framework?
It is the operating model that decides who owns AI risk in an enterprise: a model and use-case inventory, an intake gate before anything goes live, tier-appropriate validation and monitoring standards, a documented audit trail, and the committee structure that holds all of it accountable to the board or a regulator.
Is an artificial intelligence governance framework the same as an AI governance framework?
Yes. The two names mean the same thing: the roles, rules and records an organization uses to control its AI systems. The section above lists what it covers: accountability, risk, data and evidence.
What are the main AI governance frameworks?
This guide compares five published documents: the NIST AI Risk Management Framework, the ISO/IEC 42001 management system standard, the EU AI Act, the OECD AI Principles and Singapore's Model AI Governance Framework. They differ in kind: NIST describes its framework as voluntary, ISO/IEC 42001 is a standard for an AI management system, the EU AI Act is a regulation, and the OECD text is a set of principles. A framework you write yourself can borrow from more than one.
What is the difference between an AI governance framework, model and standard?
A standard, such as ISO/IEC 42001 for AI management systems, is a formal document issued by a standards body. A published framework, such as the NIST AI RMF, gives structure and vocabulary for managing AI risk. In this guide, an AI governance model means your own operating model: who decides, at what level, on what evidence. Your own framework is the document you write for your systems. It can draw on all three.
What is the Model AI Governance Framework?
It is guidance published by Singapore's Personal Data Protection Commission, first in January 2019 with a second edition in January 2020. It rests on two guiding principles, that AI decisions should be explainable, transparent and fair, and that AI systems should be human-centric. It organizes practice into four areas: internal governance, the level of human involvement in AI-augmented decisions, operations management, and stakeholder communication.
Is there an AI governance framework PDF I can download?
The primary documents come from their publishers: NIST publishes AI RMF 1.0 on its AI RMF page, and Singapore's PDPC publishes the second edition of its Model AI Governance Framework. ISO/IEC 42001 is available from ISO.
What is an AI governance playbook?
A playbook is the operational half of a framework: the step-by-step procedures for intake, testing, approval, monitoring and incidents that the team follows day to day. NIST publishes a companion AI RMF Playbook alongside its framework. Your own playbook should be specific to your systems, owners and tools, and should point back to the framework section each step serves.
How is an AI governance framework different from model risk management under SR 26-2?
SR 26-2, issued in April 2026 by the Federal Reserve, OCC and FDIC, is model risk management guidance for banking organizations and replaces SR 11-7; its cover letter says it is expected to be most relevant to banking organizations with over 30 billion dollars in total assets regulated by the Federal Reserve. It says generative and agentic AI models are not within its scope. An AI governance framework applies the same discipline of inventory, validation, monitoring and accountability to all AI systems, including generative and agentic ones, in any sector. A bank's framework should sit alongside its model risk program, not duplicate it.
Does an AI governance framework need to be different for generative and agentic AI?
The governance structure, inventory, intake, committee, and audit trail, stays the same. What changes is validation: a generative or agentic system is tested with structured output sampling against a rubric, red-teaming for guardrail failures, and adversarial testing of the tools an agent can call, rather than the statistical backtesting used for a traditional predictive model.
Talk to the team that would do the work
Bring your requirements to a working session with the person who'll actually deliver.
Book a Discovery Call