Skip to main content

NIST AI RMF — AI Risk Management in AIDLC

12 min read
AI documentation guide

Overview​

The NIST AI Risk Management Framework (AI RMF) is a voluntary framework for managing risks throughout the design, development, deployment, and use of AI systems. This guide maps AIDLC activities to AI RMF 1.0 (NIST AI 100-1, January 2023) and the Generative AI Profile (NIST AI 600-1, July 2024).

  • Scope: The framework is not limited to a particular industry or federal procurement. Select activities according to the system's context, risk tolerance, and organizational responsibilities.
  • Legal requirements: Distinguish the voluntary framework from obligations in applicable laws, agency policies, and contracts. An AI RMF mapping alone does not establish compliance.
  • Evidence: Connect each risk to an owner, evaluation method, decision record, and residual risk.

4 Functions — GOVERN, MAP, MEASURE, MANAGE​

GOVERN applies across the other three functions. MAP, MEASURE, and MANAGE are activities repeated throughout the system lifecycle, rather than a fixed certification sequence.

Figure 1. The four NIST AI RMF functions (GOVERN, MAP, MEASURE, MANAGE) and how they relate

Expand the diagram to read labels and follow connections.
The four NIST AI RMF functions
GOVERN supports MAP, MEASURE, and MANAGE as they repeat context identification, measurement, and risk response.

The identifiers below are selected official AI RMF Core subcategories. Their descriptions are summaries; the AIDLC mappings are this guide's implementation examples.

1. GOVERN​

Purpose: Establish the policies, responsibilities, and organizational basis for risk management.

  • GOVERN 1.1: Understand, manage, and document applicable legal and regulatory requirements.
  • GOVERN 1.2: Incorporate trustworthy AI characteristics into organizational policies, processes, and practices.
  • GOVERN 1.3: Determine the level of risk management needed according to organizational risk tolerance.
  • GOVERN 2.1: Document clear roles, responsibilities, and communication paths.

AIDLC mapping: Connect these activities to owners, review criteria, and exception records in the governance framework.

2. MAP​

Purpose: Understand the system's context and potential impacts.

  • MAP 1.1: Document intended uses, users, deployment contexts, expectations, and limitations.
  • MAP 1.6: Elicit and understand system requirements with relevant stakeholders.
  • MAP 2.1: Define the tasks and implementation methods the system will support.
  • MAP 4.1: Identify technical and legal risks in components, including third-party data and software.
  • MAP 5.1: Document the likelihood and magnitude of identified impacts.

AIDLC mapping: Specify context, dependencies, and failure impacts during Inception's Requirements Analysis and Reverse Engineering.

3. MEASURE​

Purpose: Evaluate identified risks and trustworthiness characteristics, including measurement limitations.

  • MEASURE 1.1: Select methods for priority risks and document risks that cannot be measured.
  • MEASURE 2.1: Document evaluation conditions, including test data, metrics, and tools.
  • MEASURE 2.3: Evaluate performance or assurance criteria under conditions similar to deployment.
  • MEASURE 2.7: Evaluate security and resilience.
  • MEASURE 2.9 / 2.10 / 2.11: Evaluate explanation and interpretation, privacy risk, and fairness and bias, respectively.

AIDLC mapping: Connect these activities to Construction's Build & Test and harness engineering quality gates. Code coverage and successful SAST scans do not substitute for evaluating AI fairness or reliability.

4. MANAGE​

Purpose: Allocate risk-response resources according to evaluation results and improve operations.

  • MANAGE 1.1: Decide whether the system meets its objectives and whether development or deployment should proceed.
  • MANAGE 1.2: Prioritize risk treatment by impact, likelihood, and available resources.
  • MANAGE 2.3 / 2.4: Respond to and recover from new risks; disengage or replace the system when needed.
  • MANAGE 3.1: Monitor risks and benefits of third-party resources.
  • MANAGE 4.1 / 4.3: Implement monitoring, change, and recovery plans and communicate incidents to relevant parties.

AIDLC mapping: Connect these activities to monitoring, incident response, rollback, and risk reassessment during Operations.

AI RMF 1.0 and the Generative AI Profile​

Table 1. Role and use of AI RMF 1.0, 600-1, and the Playbook

ResourceRoleUse
AI RMF 1.0, January 2023Defines the four functions, categories, and subcategoriesDesign system-specific risk activities and evidence records
NIST AI 600-1, July 2024A cross-sectoral generative AI profile of AI RMF 1.0Select generative AI risk actions for the intended context
AI RMF PlaybookCompanion implementation resourcesConsult suggested activities and review questions

The Generative AI Profile is not a separate AI RMF 1.1 revision. This guide uses the publications identified above; check NIST's publication history when adopting the framework.

U.S. Federal Procurement and Policy History​

EO 14110 was issued on October 30, 2023, but EO 14148 revoked it on January 20, 2025. It should therefore not be cited as a current blanket requirement for all federal contractors to comply with AI RMF, or as a current model-training reporting obligation. The status of individual regulations and agency actions after revocation must be checked separately.

OMB M-25-21 and M-25-22, issued in April 2025, address federal AI use and acquisition. This guide does not determine their applicability to a particular contract. Review the OMB guidance, agency policy, solicitation, and contract clauses applicable at the time, then map the required controls to AI RMF activities.

Record: The applicable document and version, agency and contract scope, requirement identifier, control owner, evidence location, and review date.

AIDLC Integration Examples​

The YAML below illustrates project-owned records. It is neither a NIST standard schema nor configuration automatically executed by AIDLC tooling. Validate and choose thresholds and tests for the specific use case.

Inception Stage: GOVERN + MAP​

project: federal-contract-ai-tool
assessment_date: "2026-09-18"
framework: "NIST AI RMF 1.0"
governance:
  references: ["GOVERN 1.1", "GOVERN 2.1"]
  responsible_team: "AI Governance Team"
  applicable_requirements: [] # Populate from the applicable contract and policy review.
context:
  references: ["MAP 1.1", "MAP 2.1", "MAP 4.1"]
  intended_use: "Generate draft backend code for human review"
  autonomous_deployment: false
  dependencies: ["model-provider", "code-repository"]
risks:
  - id: RISK-001
    description: "Generated code introduces vulnerabilities"
    planned_controls: ["security testing", "human review"]
  - id: RISK-002
    description: "Prompts disclose restricted information"
    planned_controls: ["data minimization", "access controls"]

Construction Stage: MEASURE​

evaluation_plan:
  references: ["MEASURE 1.1", "MEASURE 2.1", "MEASURE 2.3"]
  dataset_revision: "replace-with-reviewed-revision"
  model_revision: "replace-with-pinned-revision"
  task_success:
    minimum_rate: 0.95 # Illustrative project threshold, not a NIST requirement.
    uncertainty: "Record sample size and a confidence interval"
  security:
    reference: "MEASURE 2.7"
    checks: ["SAST", "dependency review", "prompt injection tests"]
  privacy:
    reference: "MEASURE 2.10"
    checks: ["prompt data review", "output disclosure tests"]
  fairness:
    reference: "MEASURE 2.11"
    decision: "Define relevant groups and metrics, or document non-applicability"
  unmeasured_risks: [] # Record limitations before a release decision.

Operations Stage: MANAGE​

operations_plan:
  references: ["MANAGE 2.4", "MANAGE 4.1", "MANAGE 4.3"]
  task_failure_rate:
    unit: "fraction of evaluated tasks"
    observation_window: "1h"
    warning_threshold: 0.008 # 0.8%; illustrative.
    release_limit: 0.01 # 1%; illustrative.
    minimum_evaluated_tasks: 1000 # Choose using a workload-specific sample plan.
    missing_evaluations: "Track coverage separately; do not treat as success"
  incident_response:
    owner: "on-call platform team"
    actions: ["pause automated actions", "review impact", "restore approved revision"]
    communication: "Follow the applicable incident notification plan"
  risk_review:
    triggers: ["model change", "material incident", "new use case", "scheduled review"]

References​

Official framework:

Policy history and procurement:

Related documents: