A risk committee is asked to approve a customer-facing assistant. Someone asks which standard the review followed. For a credit model the answer is familiar. For a system that writes its own answers, many teams reach for the NIST AI Risk Management Framework.
The framework is not a regulation, and no examiner enforces it. It is useful for a narrower reason. It gives business, risk and compliance teams one structure for the same questions, and NIST has published a companion profile that applies that structure to generative AI.
- The NIST AI RMF (AI 100-1), released January 26, 2023, is voluntary and sets no thresholds. The firm sets its own risk tolerance. As of November 11, 2026, AI RMF 1.0 is still the current version and is being revised, with no date given.
- The Generative AI Profile (AI 600-1), released July 26, 2024, defines twelve risks. Confabulation, data privacy, harmful bias and information integrity show up in almost every customer-facing review.
- Regulated firms use it because other rules point to it, it sits on top of existing risk practice, and it fills the gap SR 26-2 leaves for generative and agentic AI.
- The four functions, Govern, Map, Measure and Manage, are not a checklist or an ordered set of steps. The framework describes outcomes, and a reviewer needs records for each one.
- The AI RMF does not replace a sector rule or meet an EU AI Act obligation. It gives a structure for the evidence those rules ask for.
What the AI RMF is, and what "voluntary" means
NIST released the AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1) on January 26, 2023. NIST says it "is intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems" (NIST AI RMF page).
The document describes itself the same way: "voluntary, rights-preserving, non-sector-specific, and use-case agnostic" (AI 100-1, Executive Summary). It sets no thresholds. "While the AI RMF can be used to prioritize risk, it does not prescribe risk tolerance" (Section 1.2.2). The firm sets its own tolerance, and the same section says organizations "should follow existing regulations and guidelines" where those exist.
On July 26, 2024, NIST released the Generative AI Profile (NIST AI 600-1). It is "a cross-sectoral profile of and companion resource for the AI Risk Management Framework (AI RMF 1.0) for Generative AI" (AI 600-1, Section 1). It defines the risks that are "novel to or exacerbated by the use of GAI" and lists suggested actions under each function. The AI RMF Playbook adds suggested actions for every subcategory and says it "is neither a checklist nor set of steps to be followed in its entirety."
Status as of November 11, 2026. AI RMF 1.0 is still the current version. The NIST AI RMF page states: "The AI RMF 1.0 is being revised as part of the White House AI Action Plan." It gives no date. The Playbook page says the Playbook "will be updated after the AI RMF is revised." The page lists one published profile, AI 600-1, and a concept note released on April 7, 2026 for a profile on trustworthy AI in critical infrastructure. It lists no profile for AI agents. Check the page before you cite a version.
Why regulated firms use it anyway
Three reasons come up in practice. The first two are in primary sources. The third is our reading.
- Other rules point to it. The NAIC AI Model Bulletin says an insurer's AIS Program "may adopt, incorporate, or rely upon, in whole or in part, a framework or standards developed by an official third-party standard organization, such as the National Institute of Standards and Technology (NIST) Artificial Intelligence Risk Management Framework, Version 1.0" (NAIC bulletin, guideline 1.5).
- It is built to sit on top of existing risk practice. "The Framework is intended to be flexible and to augment existing risk practices which should align with applicable laws, regulations, and norms" (AI 100-1, Section 1.2.2). A bank or insurer does not have to replace its model risk or compliance process to use it.
- It fills a gap. SR 26-2 puts generative and agentic AI outside its scope and says the bank's own "risk management and governance practices should guide" the controls (SR 26-2 attachment, footnote 3). A bank that has to write its own standard needs a starting structure. In our view the AI RMF is a reasonable one, because it is public, free and already written for generative AI.
The seven trustworthiness characteristics
The framework defines what a trustworthy AI system looks like before it says how to manage risk. The characteristics are "valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed" (AI 100-1, Section 3).
Two points matter for a review team. First, "Valid & Reliable is a necessary condition of trustworthiness" (Figure 4). A system that gives wrong answers cannot be made trustworthy by adding disclosures. Second, the numbers are a human decision: "Human judgment should be employed when deciding on the specific metrics related to AI trustworthiness characteristics and the precise threshold values for those metrics" (Section 3). The framework will not tell you what pass rate is good enough.
The generative AI risks named in AI 600-1
Section 2 of the Generative AI Profile defines twelve risks. Four of them show up in almost every customer-facing review.
- Confabulation. "The production of confidently stated but erroneous or false content (known colloquially as 'hallucinations' or 'fabrications') by which users may be misled or deceived." The Profile adds that this "may be especially important to monitor when integrating GAI into applications involving consequential decision making" (Section 2.2).
- Data Privacy. "Impacts due to leakage and unauthorized use, disclosure, or de-anonymization of biometric, health, location, or other personally identifiable information or sensitive data."
- Harmful Bias or Homogenization. The definition includes "performance disparities between sub-groups or languages." An assistant that serves one group of customers worse than another falls here (Section 2.6).
- Information Integrity. A "lowered barrier to entry to generate and support the exchange and consumption of content which may not distinguish fact from opinion or fiction or acknowledge uncertainties."
The other eight are CBRN Information or Capabilities; Dangerous, Violent, or Hateful Content; Environmental Impacts; Human-AI Configuration; Information Security; Intellectual Property; Obscene, Degrading, and/or Abusive Content; and Value Chain and Component Integration. Two of these deserve a second look in a regulated firm. Human-AI Configuration covers "automation bias" and "over-reliance" by the people who use the system. Value Chain and Component Integration covers "non-transparent or untraceable integration of upstream third-party components," which describes most applications built on a vendor's foundation model.
The Profile has limits, and it states them. It covers risks with "an existing empirical evidence base at the time this profile was written." "Not every subcategory of the AI RMF is included in this document." And "suggested actions relevant to GAI developers may not be relevant to GAI deployers" (AI 600-1, Sections 2 and 3). A firm that deploys a vendor model has to select the actions that fit its role.
The twelve risks do not include a separate category for agents. The phrase "autonomous agents" appears once, in a list of security threats (MS-2.7-001). Our inference: for an agent, the Map step also has to record which tools it can call and which actions it can take without a person, and the tests have to cover those actions, not only the text it writes.
The four functions, applied to a GenAI application
Section 5 of AI 100-1 organizes the work into four functions. It is clear about what they are not: "Actions do not constitute a checklist, nor are they necessarily an ordered set of steps."
Govern
Govern is "a cross-cutting function that is infused throughout AI risk management and enables the other functions of the process" (Section 5.1). It asks for policies, an inventory (GOVERN 1.6), and named accountability: "Executive leadership of the organization takes responsibility for decisions about risks associated with AI system development and deployment" (GOVERN 2.3). For generative AI, the Profile suggests the firm "establish minimum thresholds for performance or assurance criteria and review as part of deployment approval" policies (GV-1.3-002).
Map
Map "establishes the context to frame risks related to an AI system" (Section 5.2). The output is a decision: enough "contextual knowledge about AI system impacts to inform an initial go/no-go decision." For a customer-facing assistant, that means writing down who will use it, what it may and may not answer, which data it retrieves, and which of the twelve risks apply. The Profile asks teams to consider "internal vs. external use, narrow vs. broad application scope, fine-tuning, and varieties of data sources (e.g., grounding, retrieval-augmented generation)" (MP-1.1-001).
Measure
Measure "employs quantitative, qualitative, or mixed-method tools, techniques, and methodologies to analyze, assess, benchmark, and monitor AI risk and related impacts." One sentence sets the cadence: "AI systems should be tested before their deployment and regularly while in operation" (Section 5.3). The framework also asks who does the testing. "Internal experts who did not serve as front-line developers for the system and/or independent assessors are involved in regular assessments," and domain experts and users are consulted (MEASURE 1.3).
The Profile is direct about weak testing. Teams should "avoid extrapolating GAI system performance or capabilities from narrow, non-systematic, and anecdotal assessments" (MS-2.5-001). It warns that current pre-deployment tests "may be inadequate, non-systematically applied, or fail to reflect or mismatched to deployment contexts" (AI 600-1, Appendix A.1.4). A demo that went well is not a measurement.
Manage
Manage "entails allocating risk resources to mapped and measured risks on a regular basis and as defined by the GOVERN function" (Section 5.4). It starts with the approval itself: "A determination is made as to whether the AI system achieves its intended purposes and stated objectives and whether its development or deployment should proceed" (MANAGE 1.1). It continues after launch, with post-deployment monitoring plans (MANAGE 4.1) and a way "to supersede, disengage, or deactivate AI systems that demonstrate performance or outcomes inconsistent with intended use" (MANAGE 2.4).
The vendor model is in scope here too. "Pre-trained models which are used for development are monitored as part of AI system regular monitoring and maintenance" (MANAGE 3.2). The Profile suggests teams "re-assess model risks after fine-tuning or retrieval-augmented generation implementation" and for use cases "that were not evaluated in initial testing" (MG-3.1-003).
What evidence each function produces
The framework describes outcomes. A reviewer needs records. The table pairs each function with a suggested action from the Generative AI Profile and the evidence a team would keep for a customer-facing application or agent. The evidence column is our reading, not NIST's.
| Function | What NIST asks | Evidence for a GenAI application |
|---|---|---|
| Govern | "Enumerate organizational GAI systems for incorporation into AI system inventory" (GV-1.6-001). Set approval thresholds (GV-1.3-002). | An inventory entry with an owner. A written risk tolerance. The thresholds the application must pass, and who can approve it. |
| Map | Document intended purpose and context of use (MAP 1.1). "Organizational risk tolerances are determined and documented" (MAP 1.5). | The use case, the users, the data sources, the actions an agent can take, and the list of risks that apply to it. |
| Measure | "Review and verify sources and citations in GAI system outputs during pre-deployment risk measurement and ongoing monitoring activities" (MS-2.5-003). "Share results of pre-deployment testing with relevant GAI Actors, such as those with system release approval authority" (MS-2.3-003). | Test sets and results for each mapped risk. Ratings from domain experts who did not build the system. A report the approver has read. |
| Manage | "Establish, maintain, and evaluate effectiveness of organizational processes and procedures for post-deployment monitoring of GAI systems" (MG-4.1-002). Policies and procedures "to record and track GAI system reported errors, near-misses, and negative impacts" (MG-4.3-002). | The approval decision and its basis. A log of issues from raised to resolved. Monitoring results, and the criteria that trigger a retest or a shutdown. |
How the AI RMF relates to other regimes
The AI RMF does not replace a sector rule. It gives a structure for the evidence that sector rules ask for.
- US banking. SR 26-2 leaves generative and agentic AI to the bank's own practices. Our guide to SR 26-2 and generative AI covers what model risk teams still owe examiners. The four functions are one way to organize that work.
- US insurance. The NAIC bulletin names the AI RMF as a framework an AIS Program may rely on (guideline 1.5). See our guide to the NAIC AI Model Bulletin.
- ISO/IEC 42001. NIST's AI Resource Center publishes a "NIST AI Risk Management Framework to ISO-IEC-42001 Crosswalk" (AIRC crosswalks). The crosswalk shows where the two line up. See our guide to ISO 42001 and generative AI.
- EU AI Act. The Act is law, and the AI RMF is not. Following the AI RMF does not meet an EU obligation. The same records can still support both. See our guide to the EU AI Act and generative AI.
How GenGuardX helps
GenGuardX is a GenAI governance platform for regulated teams. Business teams test an AI application, risk teams approve it against evidence, and both monitor it after launch. The table shows how that evidence can be organised against the four functions. The mapping supports a program built on the AI RMF. It does not make an organization compliant with the framework or with any regulation.
| Function | GenGuardX capability | What is recorded |
|---|---|---|
| Govern | Inventory management, approval workflows and role-based access in the Governed Workbench. | Each prompt, model, RAG and agent with an owner. Who approved the application, on what evidence, and when. |
| Map | A risk library to identify the risks that apply, and lineage tracking. | The risks selected for the use case, and the datasets, prompts and models the application depends on. |
| Measure | Standardized evaluations, Human Integrated Testing, annotation queues, and Risk dashboards & curated reports. | Versioned, re-runnable test results for each risk. Business users run realistic scenarios, flag issues and rate responses. |
| Manage | A findings database, production monitoring & alerts, and re-evaluation. | Every issue tracked from raised to resolved. Drift from the approved version. Production findings that feed the next round of tests. |
The point of the mapping is a shared record. The business owner, the risk reviewer and the examiner can each find the same answer to the same four questions: what the system is for, how it was tested, who approved it, and how the firm knows it still works.
FAQ
Is the NIST AI RMF mandatory?
No. NIST says the framework "is intended for voluntary use." The Playbook adds that its "suggestions are voluntary." Other rules can still point to it. The NAIC AI Model Bulletin, for example, says an insurer's AIS Program may rely on the NIST AI RMF (guideline 1.5).
What are the four functions of the NIST AI RMF?
Govern, Map, Measure and Manage. Govern is "a cross-cutting function" that runs through the other three. Map sets the context and identifies risks. Measure tests and monitors them. Manage decides what to do about them, before and after launch (AI 100-1, Section 5).
What is NIST AI 600-1?
NIST AI 600-1 is the Generative AI Profile, released on July 26, 2024. It is "a cross-sectoral profile of and companion resource for" the AI RMF. It defines twelve risks that are novel to or exacerbated by generative AI and lists suggested actions under the four functions.
Has NIST updated the AI RMF or published a profile for AI agents?
As of November 11, 2026, AI RMF 1.0 is still the current version. NIST's AI RMF page says it "is being revised as part of the White House AI Action Plan." The page lists one published profile, the Generative AI Profile, and an April 7, 2026 concept note for a critical infrastructure profile. It lists no profile for AI agents.
Last reviewed: November 11, 2026. This article summarizes public guidance and is not legal advice.
See how GenGuardX organises GenAI approval evidence against the NIST AI RMF.