NIST AI RMF & AI Compliance

Overview

The NIST AI Risk Management Framework (AI RMF 1.0), released by the National Institute of Standards and Technology in January 2023, is a voluntary framework designed to help organizations manage risks associated with AI systems throughout their lifecycle. The AI RMF is a flexible, structured approach that organizations can adapt to their specific context, capabilities, and risk tolerance. It has become a leading reference framework for AI governance in the United States and is increasingly recognized internationally. NIST's Generative AI Profile (NIST AI 600-1, July 2024) extends it to generative models.

The framework is organized into two main sections: the foundational information that frames AI risks, and the Core functions that provide actionable guidance. The Core consists of four functions: Govern, Map, Measure, and Manage. Each contains categories and subcategories that describe specific outcomes and activities. This structure is intentionally aligned with the NIST Cybersecurity Framework's organize-by-function approach, making it accessible to organizations already experienced with NIST standards.

For AI development teams, the AI RMF provides a common language and systematic methodology for identifying, assessing, and mitigating AI risks. It addresses risks related to bias and fairness, transparency and explainability, privacy, security, safety, and accountability. The framework treats AI risk management as part of broader organizational risk management, so that AI-specific risks are considered alongside operational, financial, and reputational risks.

AI-Specific Requirements

The Govern function establishes the organizational structures, policies, and processes for AI risk management. It calls for cultivating a risk-aware culture, defining roles and responsibilities for AI governance, implementing policies and procedures for AI development and deployment, and establishing mechanisms for ongoing monitoring and review. Organizations must document their AI risk tolerance, define escalation procedures for identified risks, and back AI governance with adequate resources and executive commitment.

The Map function focuses on understanding the context in which AI systems operate. This includes identifying and categorizing AI systems, assessing the potential impacts of AI system failures or misuse, understanding the characteristics of the data used for training and operation, and evaluating the operational environment and stakeholder expectations. For training data specifically, the Map function calls for understanding data provenance, assessing data quality and representativeness, identifying potential biases, and documenting data collection and processing methodologies.

The Measure function addresses the assessment and analysis of AI risks through quantitative and qualitative methods. Organizations should implement testing and evaluation procedures, measure AI system performance against defined metrics, assess bias and fairness across relevant demographic groups, evaluate security and privacy risks, and conduct regular reassessments as systems and contexts evolve. The Manage function then takes the outputs of mapping and measurement to prioritize and treat identified risks, implement controls, communicate risk information to stakeholders, and monitor the effectiveness of risk treatments.

How Ertas Helps

Several of the risks the framework asks you to map come from where data and models live. In Ertas Studio, training runs on Ertas's cloud GPUs, so the training dataset is uploaded to Ertas and stored so runs can be repeated, and you can delete it when it is no longer needed. Record that flow in your Map work, and reduce privacy risk by training on de-identified or minimized data, with each example edited and approved in Data Craft, Studio's dataset builder.

For Measure, Studio lets you run several training runs in parallel and compare them, which gives you evidence for choosing one candidate model over another. The tests that matter in your context, such as bias, robustness and task accuracy, are yours to design. For Manage, the trained model is yours: export it as GGUF or safetensors and run it in your own environment, where inference data never reaches Ertas. Its version stays fixed until you retrain, so the model you measured is the model in production. The adaptive model endpoint and enterprise builds can run the model in your own cloud or on-premise.

Govern stays with you entirely: policies, roles, risk tolerance and stakeholder engagement. So do the risk register and the impact assessments that feed it.

Compliance Checklist