The NIST AI Risk Management Framework is the most widely referenced AI governance standard in the United States. It is voluntary. Federal agencies, large customers, and increasingly the auditors who go in behind them all reference it as if it were not. That tension (voluntary in the statute, expected in practice) is where most NIST AI RMF compliance engagements actually start.
I sat inside the U.S. FDIC in 2021 as federal agencies began reckoning with the draft framework. The firms that have done well with it since then have not treated it as a standard to conform to. They have treated it as a diagnostic to run against themselves, a quarterly exercise in honest self-assessment, and the basis of an operating program that draws on the framework’s vocabulary without becoming a separate NIST workstream layered on top of existing risk processes. That is the model Virtova brings into engagements.
Detailed thinking on operationalizing the framework, including the July 2024 Generative AI Profile (AI 600-1), sits in the longer Virtova NIST AI RMF playbook. The service page below covers what a compliance engagement looks like in practice.
What this engagement looks like
A Virtova NIST AI RMF compliance engagement typically runs six to ten weeks for the diagnostic, with optional follow-on remediation. The work is organized around the framework’s four functions plus a fifth integration thread.
Govern. The thinnest function in nearly every engagement. We build the operating muscle: a named senior accountable executive, a working risk forum, a documented risk appetite statement that says specific things, and a decision log that survives the next exam. Govern is what turns the rest of the framework from documentation into program.
Map. The function that asks the firm to understand each AI system in context: data sources, downstream consumers, the human decisions it informs, the failure modes that matter, the affected parties, and (for generative systems under AI 600-1) training-data provenance and confabulation surface. A good Map artifact for a single system fits on two pages, in plain English. If a senior examiner cannot follow it, it is not yet a Map.
Measure. Where most firms bring model-quality metrics (accuracy, F1, ROC-AUC) to what is supposed to be a risk conversation. We translate model behavior into a measurement set the board and supervisors will actually use: tier-appropriate validation evidence, drift monitoring, fairness and disparate-impact metrics for in-scope use cases, and incident tracking. Measure is also where the gap between U.S. and EU expectations starts to bite for firms with cross-border exposure.
Manage. The closing-the-loop function. Mitigation actions tied to specific findings, change-management discipline for model updates and vendor swaps, retirement criteria, and the documented response when measures cross a threshold. Manage is where the program proves it can act, not just observe.
Integration. The thread the framework itself does not name explicitly: how AI risk management connects to the firm’s existing model risk management, third-party risk, information security, fair-lending, and BSA/AML programs. Treating AI risk as a separate vertical is the most common implementation mistake we see. Virtova engagements wire the four NIST functions into existing risk infrastructure rather than building a parallel one.
When the engagement is the wrong answer
NIST AI RMF compliance work is the wrong scope when the firm needs a foundational AI governance program built end-to-end. At that point the right engagement is AI governance consulting, with NIST RMF as one input among several. It is also the wrong scope when the firm has a working program and the ask is specifically about a single function (Map remediation after an audit finding, for example), at which point a tightly bounded sprint fits better than the diagnostic.
Virtova will tell you which one fits on the discovery call.
Generative AI: the AI 600-1 layer
The July 2024 NIST Generative AI Profile (AI 600-1) extends the RMF with specific risks and controls for generative systems: confabulation, data integrity, environmental and human-rights considerations, harmful bias, intellectual-property exposure, and obscenity. For most U.S. firms in 2026, the Generative AI Profile is the more operationally relevant document day to day; the base RMF is the structural anchor underneath it. Engagements scope both.
In practice, the most common AI 600-1 gaps Virtova finds are around training-data provenance documentation, evaluation evidence for confabulation and harmful-bias risks at the use-case level, and the human-oversight design for generative systems that have effectively been shipped without one. Closing those three gaps tends to be the largest single piece of remediation work in a typical engagement.
Next step
Most engagements start with a 30-minute discovery call. Bring whatever you have (current charter, model inventory, last audit finding, the AI section of the most recent board report) and we will tell you which function has the largest gap and what a tightly scoped engagement looks like.