EAAPLEnterprise AI Architecture Pattern Library
EAAPL›Library›Regulatory Compliance›
Proven
⇄ Compare

EAAPL-CMP006 — NIST AI Risk Management Framework Implementation

EAAPL-CMP006 — NIST AI Risk Management Framework Implementation

Status: Proven
Tags: nist-ai-rmf model-risk governance traceability high-complexity
Version: 1.2
Last Updated: 2026-06-12
Author: Enterprise AI Architecture Pattern Library

1. Executive Summary

The NIST AI Risk Management Framework 1.0 (NIST AI RMF, published January 2023) provides a voluntary, non-prescriptive framework for organisations to identify, assess, and manage risks throughout the AI lifecycle. Unlike ISO 42001 (certifiable) or the EU AI Act (legally binding), NIST AI RMF is a risk-based practice guide designed for broad adoption across industries and organisation sizes. Its four core functions—GOVERN, MAP, MEASURE, and MANAGE—provide a structured vocabulary and set of practices that complement any regulatory compliance programme.

This pattern provides an enterprise architecture for implementing all four NIST AI RMF core functions, mapping each function's actions (from GOVERN 1.1 through MANAGE 4.4) to specific architectural controls, governance processes, and technology components. It enables organisations to demonstrate structured AI risk management to US federal agencies, financial regulators (OCC Model Risk Management guidance aligns with RMF principles), and enterprise customers requiring evidence of AI governance maturity. The pattern is designed to integrate with ISO 42001 AIMS (where deployed) and to serve as a standalone governance framework where ISO 42001 is not pursued.

2. Problem Statement

Business Problem

US federal contractors and regulated entities (banks under OCC SR 11-7 Model Risk Management guidance, healthcare entities under FDA AI/ML guidance) face growing expectations from regulators and customers to demonstrate structured AI risk management. Without a recognised framework, organisations cannot demonstrate governance maturity, cannot respond coherently to regulatory inquiries, and cannot consistently identify and mitigate AI risks before they materialise as incidents.

Technical Problem

AI risk management in most organisations is fragmented: security teams assess cybersecurity risk, data scientists assess model performance risk, ethics committees assess fairness risk—but there is no unified framework that connects these assessments to a common risk taxonomy, aggregates findings into a consolidated risk view, and ensures that identified risks receive appropriate treatment with tracked accountability.

Symptoms

  • AI risk assessments use different methodologies and risk taxonomies across business units
  • AI risk findings are not tracked to closure with accountability
  • There is no consolidated AI risk view available to senior management
  • AI systems are deployed without formal risk treatment plans
  • Residual AI risks are not documented or accepted by appropriate risk owners

Cost of Inaction

Dimension Consequence
Regulatory US federal contract ineligibility (NIST AI RMF referenced in EO 14110); OCC enforcement for model risk management gaps
Commercial Loss of enterprise contracts requiring AI governance evidence; cyber insurance exclusions
Operational Unmanaged AI risks materialise as incidents—bias at scale, AI system failures, data breaches
Strategic Cannot participate in US federal AI programmes that reference NIST AI RMF (e.g., AI Safety Institute evaluations)

3. Context

When to Apply

  • Organisations serving the US federal government or seeking federal AI safety evaluation
  • Financial services entities subject to OCC Model Risk Management guidance (SR 11-7 equivalent)
  • Healthcare organisations seeking FDA AI/ML Software as a Medical Device alignment
  • Any enterprise seeking a voluntary, internationally recognised AI risk management framework
  • Organisations where ISO 42001 certification is not required but structured governance is

When NOT to Apply

  • Where legally binding obligations (EU AI Act, APRA) require specific compliance architecture—NIST AI RMF complements but does not replace these obligations
  • Organisations in early AI exploration with no production AI (too early for full framework adoption)

Prerequisites

Prerequisite Description
AI Inventory Basic catalogue of AI systems and intended uses
Risk Appetite Statement Documented organisational risk tolerance that can be applied to AI risks
Executive Sponsorship Senior leadership commitment to AI risk management (GOVERN function)
Cross-Functional Team AI risk management requires collaboration between technology, legal, data science, ethics, and operations

Industry Applicability

Industry NIST AI RMF Relevance Key Alignment
US Federal Contractors Mandatory reference (EO 14110) All four functions; AI Safety evaluations
Financial Services High — OCC, Fed, FDIC model risk expectations GOVERN + MAP: model risk inventory; MEASURE: performance validation
Healthcare / Life Sciences High — FDA AI/ML SaMD guidance MAP: intended use documentation; MEASURE: clinical performance; MANAGE: adverse event
Defence / National Security High — DoD AI ethical principles GOVERN: AI ethics programme; MAP: mission risk context
Technology / AI Developers Medium-High All functions; AI product liability risk management
Higher Education and Research Medium GOVERN and MAP for AI research governance

4. Architecture Overview

The NIST AI RMF Implementation Architecture mirrors the framework's four core functions, each with a distinct architectural layer that operationalises the function's actions into enterprise processes and technical controls.

GOVERN Function — AI Risk Governance Structures The GOVERN function is the foundation for all other functions. It establishes the organisational policies, processes, and accountabilities that enable MAP, MEASURE, and MANAGE to operate consistently. Architecturally, GOVERN requires: an AI Risk Governance Framework document approved by senior leadership that defines risk appetite, escalation paths, and accountability structures (GOVERN 1.1); a cross-functional AI Risk Committee or equivalent governance body (GOVERN 1.2); AI risk management policies and procedures integrated into the organisation's risk management framework (GOVERN 1.3); established processes for teams to raise AI risks without fear of retaliation (GOVERN 1.4); documented organisational roles with AI risk responsibilities (GOVERN 2.1–2.2); mechanisms to understand the sociotechnical context in which AI operates (GOVERN 3.1–3.2); team practices that support diverse perspectives in AI risk identification (GOVERN 4.1–4.2); and accountability for AI risk outcomes at the organisational level (GOVERN 5.1–5.2) and at the supply chain level (GOVERN 6.1–6.2).

MAP Function — Context Establishment and Risk Identification The MAP function operationalises risk identification: before risks can be measured or managed, they must be discovered and characterised. MAP requires: establishing the context of the AI system's intended use and deployment environment (MAP 1.1–1.6); categorising AI risks using the RMF's taxonomy of trustworthiness characteristics (valid/reliable, safe, secure/resilient, explainable, privacy-enhanced, fair with managed bias, accountable/transparent) (MAP 2.1–2.3); identifying potential impacts on individuals, communities, and organisations from AI system failures or misuse (MAP 3.1–3.5); characterising the data and technical risks in the AI system's inputs, model, and outputs (MAP 4.1–4.2); and identifying the AI risks in the context of interdependencies with other systems and processes (MAP 5.1–5.2). Architecturally, MAP is operationalised through a structured AI Risk Identification process that produces an AI Risk Profile for each system, using the RMF's Trustworthiness Characteristics as the risk taxonomy.

MEASURE Function — Risk Analysis and Benchmarking The MEASURE function operationalises risk quantification and tracking. For each identified risk, MEASURE requires: defining appropriate metrics and methodologies for measuring the risk (MEASURE 1.1–1.3); applying measurement methods and documenting results (MEASURE 2.1–2.13); tracking changes in AI risks over time through continuous monitoring (MEASURE 3.1–3.3); and assessing plans and progress towards AI risk management goals (MEASURE 4.1). Architecturally, MEASURE requires an AI Model Monitoring and Evaluation Infrastructure: automated performance tracking dashboards measuring each AI system against its defined trustworthiness metrics; bias and fairness evaluation pipelines; robustness testing programmes; explainability assessments; and risk scoring models that aggregate metrics into a risk index per AI system.

MANAGE Function — Risk Response and Treatment The MANAGE function operationalises risk treatment: for each characterised and measured risk, MANAGE requires: allocating risk response resources and implementing treatment plans (MANAGE 1.1–1.4); identifying and managing residual risks and trade-offs (MANAGE 2.1–2.4); monitoring effectiveness of risk responses and updating treatment plans (MANAGE 3.1–3.2); and managing AI incidents and near-misses with lessons-learned integration (MANAGE 4.1–4.4). Architecturally, MANAGE requires a Risk Treatment Registry linking each identified risk to its treatment plan, responsible owner, implementation status, and residual risk acceptance record; an AI Incident Management process with RMF-aligned severity classification; and a feedback loop from incident findings to MAP (updating risk identification based on what was discovered in production).

5. Architecture Diagram

ARCHITECTURE DIAGRAM
flowchart TD subgraph Govern["GOVERN"] A[AI Policy and Roles] end subgraph Map["MAP"] B[AI Risk Profile] C[Impact Identification] end subgraph Measure["MEASURE"] D[Evaluation Pipeline] E[Risk Score Index] end subgraph Manage["MANAGE"] F[Risk Treatment Plans] G[Incident Management] end A --> B B --> C C --> D D --> E E --> F F --> G G -->|feedback| B E -->|dashboard| A style A fill:#dbeafe,stroke:#3b82f6 style B fill:#f0fdf4,stroke:#22c55e style C fill:#f0fdf4,stroke:#22c55e style D fill:#f0fdf4,stroke:#22c55e style E fill:#fef9c3,stroke:#eab308 style F fill:#d1fae5,stroke:#10b981 style G fill:#fee2e2,stroke:#ef4444

6. Components

Component Type Responsibility Technology Options Criticality
AI Risk Governance Framework Governance Define risk appetite, escalation, accountability for AI risk Policy document in DMS; Confluence; SharePoint Critical
AI Risk Committee Governance Cross-functional governance body for AI risk decisions Standing committee with charter; meeting cadence Critical
AI Risk Register Governance Catalogue all identified AI risks with RMF taxonomy; track treatment Archer, Resolver, Jira, custom Critical
AI Risk Profile Builder Governance Structured intake tool for MAP function: context + trustworthiness risk categorisation Custom form + workflow; OneTrust; Credo AI High
Metrics Definition Library Governance Per-trustworthiness-characteristic metric catalogue Internal wiki; Confluence; custom High
Bias and Fairness Evaluation Engineering Automated bias detection across demographic groups Fairlearn, AI Fairness 360, Aequitas, SageMaker Clarify High
Robustness Testing Pipeline Engineering Adversarial testing; out-of-distribution detection; model stability assessment TextAttack, CleverHans, IBM Adversarial Robustness Toolbox High
Explainability Engine Engineering SHAP/LIME explanations; feature importance; counterfactual explanations SHAP, LIME, Captum, Alibi, interpret-community High
Continuous Monitoring Platform Engineering Real-time AI performance tracking; drift detection; alert on metric breach Evidently AI, Fiddler, WhyLabs, SageMaker Monitor High
Risk Scoring Engine Governance Aggregate metrics into risk index per AI system; flag high-risk systems Custom scoring model; Credo AI; OneTrust Medium
Risk Treatment Registry Governance Link risks to treatment plans; track implementation status; document residual risk acceptance Archer, Jira, ServiceNow, custom Critical
AI Incident Manager Operations Classify, investigate, and close AI incidents; extract lessons learned PagerDuty, ServiceNow, Jira High
Executive AI Risk Dashboard Reporting Consolidated AI risk view across all systems; RMF function completion; trend Power BI, Tableau, Grafana, Looker High

7. Data Flow

Primary Flow (Full RMF Cycle for a New AI System)

Step Actor Function Action Output
1 AI Project Team GOVERN Register new AI system; assign risk owner AI system record in Risk Register
2 AI Risk Function MAP Conduct context establishment: document intended use, deployment environment, affected users Context document
3 AI Risk Function MAP Categorise risks using trustworthiness characteristics (valid/reliable, safe, fair, explainable, etc.) AI Risk Profile with categorised risks
4 AI Risk Function MAP Identify potential impacts on individuals, communities, organisation Impact register for this AI system
5 Data Science / MLOps MEASURE Define metrics for each identified risk category Metrics definition document
6 Data Science / MLOps MEASURE Execute evaluation pipeline (bias, robustness, explainability, performance) Evaluation report with metric values
7 Risk Scoring Engine MEASURE Aggregate metrics into risk index; classify system risk level Risk score; risk tier (Low/Medium/High/Critical)
8 AI Risk Function MANAGE For each risk above threshold: define treatment plan Risk Treatment Plans
9 Responsible Owner MANAGE Implement treatment plans; document residual risk Treatment implementation evidence; residual risk acceptance
10 Continuous Monitor MEASURE Monitor production performance; alert on metric breach Monitoring alerts; updated risk scores
11 AI Incident Manager MANAGE On incident: classify severity; investigate; extract lessons Incident report; lessons learned
12 AI Risk Function MAP Update AI Risk Profile based on incident lessons Updated Risk Profile

Error Flow

Step Failure Detection Recovery
MAP Context Insufficient Key use-case context missing; risks not identified Evaluation finds unexpected bias; incident in production Retroactive MAP; update Risk Profile; implement missed risk treatment
MEASURE Metric Not Defined Risk identified but no metric defined to measure it Audit of Metrics Definition Library; risk score gap Define metric; execute evaluation; update risk score
Treatment Plan Not Implemented Risk identified and planned but treatment not executed Risk register currency check; audit Escalate to AI Risk Committee; implement or formally accept residual risk
Monitoring Alert Not Actioned Production metric breach alert generated but not investigated SLO breach escalation; post-incident review Formal incident; root cause; update monitoring response process

8. Security Considerations

RMF-Aligned Security Controls

GOVERN Action Security Control Implementation
GOVERN 1.1 AI Risk Governance Framework includes AI security risk as a risk category Security represented on AI Risk Committee
GOVERN 6.1 Supply chain AI risk governance covers security of third-party AI components Vendor security assessment; SBOM for AI
MAP 4.1 Technical risks in AI inputs, model, and outputs include security threats Security threat modelling integrated into MAP
MEASURE 2.7 AI system security properties assessed against benchmarks Penetration testing; red-teaming; adversarial robustness testing
MANAGE 2.2 Security incidents treated as AI risk events under MANAGE function AI Incident Manager covers security incidents

OWASP LLM Top 10 — NIST AI RMF Function Mapping

OWASP LLM Risk RMF Function Actions Control
LLM01 Prompt Injection MAP + MEASURE MAP 2.2 (risk categorisation: security/resilience); MEASURE 2.7 (security assessment) Adversarial input testing; WAF deployment
LLM02 Insecure Output Handling MAP + MANAGE MAP 2.3 (impact identification); MANAGE 1.3 (treatment: output validation) Output validation controls; content filtering
LLM03 Training Data Poisoning MAP + MEASURE MAP 4.1 (data risks); MEASURE 2.5 (data quality assessment) Training data provenance; integrity verification
LLM04 Model Denial of Service MEASURE + MANAGE MEASURE 3.1 (tracking changes in AI risks); MANAGE 3.1 (monitoring effectiveness) Availability monitoring; rate limiting; DR
LLM05 Supply Chain Vulnerabilities GOVERN + MAP GOVERN 6.1–6.2 (supply chain governance); MAP 5.1 (interdependency risks) SBOM; vendor risk assessment; component integrity
LLM06 Sensitive Information Disclosure MAP + MANAGE MAP 3.1 (impact on individuals: privacy); MANAGE 1.2 (treatment: PII controls) Output PII filtering; training data governance
LLM07 Insecure Plugin Design MAP + MEASURE MAP 4.2 (technical risks: tool/plugin design); MEASURE 2.7 (security properties) Tool least-privilege; call whitelisting
LLM08 Excessive Agency GOVERN + MAP GOVERN 2.1 (accountability); MAP 1.5 (context: level of automation and autonomy) Scope limits; human approval gates
LLM09 Overreliance MAP + MEASURE MAP 1.6 (context: human oversight); MEASURE 2.9 (explainability assessment) Confidence scoring; explanation display; user education
LLM10 Model Theft GOVERN + MANAGE GOVERN 1.3 (policies: IP protection); MANAGE 2.3 (residual risk: model access control) Model weight access control; DLP; egress monitoring

9. Governance Considerations

RMF Function Governance

Function Governance Requirement Owner Cadence
GOVERN AI Risk Governance Framework maintained; AI Risk Committee operating; accountability RACI current Chief Risk Officer / AI Officer Annual review; event-driven update
MAP AI Risk Profile completed for each AI system before deployment; updated on material change AI Risk Function + Business Unit AI Owner At intake; on material change; annual refresh
MEASURE Metrics defined and measured for all production AI systems; risk scores current MLOps / AI Engineering Continuous (automated); quarterly review
MANAGE Risk Treatment Plans in place for all risks above risk appetite; residual risks formally accepted AI Risk Owner per system On risk identification; continuous monitoring

Governance Artefacts

Artefact Description RMF Action Retention
AI Risk Governance Framework Policy + structure for AI risk management organisation-wide GOVERN 1.1, 1.3 Current + 10 years
AI Risk Committee Charter and Minutes Mandate, membership, cadence, and decisions of AI Risk Committee GOVERN 1.2 10 years
AI Risk Profiles Per-system MAP output: context, risk categorisation, impact identification MAP 1–5 Lifetime of system + 7 years
Metrics Definition Library Approved metrics for each trustworthiness characteristic category MEASURE 1.1 Current + 7 years
Evaluation Reports MEASURE pipeline results per system per evaluation cycle MEASURE 2.1–2.13 7 years
Risk Treatment Plans Treatment plan per risk above appetite; implementation evidence MANAGE 1.1–1.4 7 years
Residual Risk Acceptance Records Formally approved acceptance of residual risk by appropriate risk owner MANAGE 2.4 10 years
AI Incident Reports Incident classification, investigation, root cause, and lessons learned MANAGE 4.1–4.4 10 years

10. Operational Considerations

Monitoring and SLOs

SLO Target RMF Action Breach Action
AI Risk Profile Coverage 100% of production AI systems with current Risk Profile MAP (all actions) Block new deployments; retroactive profiling for existing systems
Metrics Measurement Frequency All high-risk AI systems measured at least monthly; medium quarterly MEASURE 2.1–2.13 Alert MLOps; escalate to AI Risk Committee
Risk Score Currency All production AI systems with risk score updated within 90 days MEASURE 3.1 Automated re-evaluation trigger
Treatment Plan Completion >90% of risk treatments implemented within planned timeline MANAGE 1.1 Escalate to AI Risk Committee; formal extension or risk acceptance
Incident to Lessons Learned 100% of AI incidents have lessons learned integrated into MAP within 60 days MANAGE 4.4 → MAP Audit; escalate to AI Officer

Disaster Recovery

Scenario Impact Recovery
AI Risk Register Data Loss Loss of risk identification records; governance blindspot Restore from backup; conduct emergency risk review for P0 systems
Continuous Monitor Outage MEASURE 3.x tracking gap; risk changes not detected Alert AI Risk Committee; manual monitoring for critical systems; restore within 24 hours
AI Risk Committee Quorum Failure MANAGE decisions cannot be made Deputy decision-makers designated in charter; asynchronous decision process

11. Cost Considerations

Cost Drivers

Cost Driver Indicative Cost Notes
AI Risk Governance Platform USD 30,000–100,000/year Commercial (Credo AI, OneTrust) or build on Jira/Confluence
Bias and Fairness Evaluation Tooling USD 10,000–50,000/year Commercial (Fiddler, Credo AI) or open source (Fairlearn)
Robustness and Adversarial Testing USD 20,000–80,000/year Commercial tools; red-team exercises
Continuous Monitoring Platform USD 20,000–80,000/year Evidently AI, WhyLabs, Fiddler
AI Risk Management FTE USD 300,000–700,000/year 1–3 FTE dedicated to RMF programme
NIST AI RMF Training USD 5,000–20,000/year NIST AI RMF workshops; Playbook training

Indicative Cost Range

Organisation Annual NIST AI RMF Implementation Cost Notes
Small (<10 AI systems) USD 200,000–500,000 Lean programme; open-source tooling; 1 FTE
Mid-size (10–30 AI systems) USD 600,000–1,500,000 Commercial platform; dedicated team
Large (30+ systems, enterprise) USD 2,000,000–6,000,000 Enterprise-grade tooling; large team; integrated with existing ERM

12. Trade-Off Analysis

Implementation Approach Options

Option Description Pros Cons Recommended For
Option A: Full GOVERN + MAP + MEASURE + MANAGE Implement all four functions with all actions for all AI systems Most comprehensive; strongest evidence for regulators Highest cost and complexity; risk of AIMS becoming a paperwork burden Large enterprises; US federal contractors; regulated industries
Option B: Risk-Tiered Implementation Full implementation for high-risk AI; lighter-weight for low-risk Proportionate; reduces overhead for non-critical AI Requires consistent risk tiering methodology Most organisations; start here and expand
Option C: GOVERN + MAP First (Phase 1) Implement governance and risk identification before measurement and management Builds foundation; immediate value; lower initial cost Delayed measurement and treatment capability Organisations starting their AI risk management programme

Architectural Tensions

Tension Trade-Off Resolution
RMF Completeness vs Velocity Thorough MAP + MEASURE can add weeks to AI deployment timelines Proportionate process: lighter MAP for low-risk AI; full process for high-risk
Quantitative vs Qualitative Measurement MEASURE function ideally uses quantitative metrics; some AI risks resist quantification Hybrid approach: quantitative where metrics exist; qualitative structured assessment where not
Centralised vs Distributed Risk Ownership GOVERN centralises AI risk management; MANAGE requires local risk owners Central AI Risk Function for framework; distributed AI Risk Owners for system-level accountability
RMF as Regulatory Compliance vs Risk Management Some organisations treat RMF as a compliance tick-box GOVERN 1.4: establish culture of genuine risk identification; reward surfacing risks

13. Failure Modes

Failure Likelihood Impact Detection Recovery
GOVERN Not Implemented — MAP/MEASURE/MANAGE Operate Without Governance High Critical — no accountability for risk decisions; inconsistent implementation Audit; regulator review Establish GOVERN before other functions; retroactively assign accountability
MAP Treats All AI as Low-Risk High High — high-risk systems escape rigorous evaluation MEASURE finds unexpected issues; incident Re-classify using objective risk tiering criteria; retroactive MAP for upgraded systems
MEASURE Metrics Are Vanity Metrics Medium High — metrics show good results but miss real risks Incident exposes unmeasured risk Review metrics against risk taxonomy; add materially relevant metrics
Treatment Plans Not Funded Medium High — risks identified but not treated Risk Treatment Registry currency check Escalate funding requirement to AI Risk Committee; risk acceptance if unfunded
Incident Lessons Not Fed Back to MAP High High — same risks recur; MAP remains static Repeat incidents of same type; audit Enforce MANAGE 4.4 → MAP feedback loop; process gate

Cascading Failure Scenario

An organisation implements MEASURE before GOVERN. Individual model teams define their own metrics (accuracy, AUC) without reference to the RMF trustworthiness characteristics. No fairness or explainability metrics are defined because these were not in the initial MEASURE scope. High-risk AI systems in credit decisioning and HR are deployed with strong performance metrics but no bias assessment. A regulatory review triggered by a customer complaint discovers systematic demographic bias in credit decisions. Because there is no AI Risk Register (MAP not implemented), the organisation cannot demonstrate that bias was identified as a risk. Because GOVERN was not implemented, there is no accountability structure—no one is responsible. The absence of GOVERN means there is no policy requiring bias assessment. The regulators find negligence in the governance structure, not just a technical failure.

14. Regulatory Considerations

Regulation NIST AI RMF Alignment Architectural Control Reference
EO 14110 (US) — Safe, Secure, Trustworthy AI Directs federal agencies to adopt NIST AI RMF; requires AI risk management for federal contractors All four functions; GOVERN 1.1 EO 14110 Section 4.1
OCC SR 11-7 Model Risk Management GOVERN + MAP = model risk identification; MEASURE = model validation; MANAGE = model risk response Model risk classification; validation programme OCC SR 11-7
EU AI Act Article 9 — Risk Management MAP function = risk identification obligation; MEASURE = monitoring; MANAGE = risk treatment AI Risk Profile; Risk Treatment Plans EU AI Act Article 9
ISO/IEC 42001:2023 GOVERN ≈ Clauses 5+6; MAP ≈ Clause 6.1; MEASURE ≈ Clause 9.1; MANAGE ≈ Clause 10 Integration architecture: AIMS + RMF ISO/IEC 42001:2023 Clause 6
APRA CPG 234 GOVERN + MAP cover AI information security risk identification AI Risk Profile includes security risks APRA CPG 234
FDA AI/ML Software as a Medical Device MANAGE 4.x incident monitoring aligns with FDA post-market safety reporting obligations AI Incident Manager with FDA report generation FDA AI/ML SaMD Action Plan
NIST Cybersecurity Framework RMF MAP aligns with CSF Identify; MEASURE with Detect; MANAGE with Respond/Recover Integration with existing CSF programme NIST CSF 2.0

15. Reference Implementations

AWS

Component AWS Service / Tool
AI Risk Register Jira on AWS; ServiceNow SaaS
Context and Risk Profiling Custom intake form on AWS Amplify
Bias and Fairness Evaluation Amazon SageMaker Clarify
Robustness Testing AWS SageMaker + custom adversarial test scripts
Explainability Amazon SageMaker Clarify (SHAP)
Continuous Monitoring Amazon SageMaker Model Monitor + CloudWatch
Risk Dashboard Amazon QuickSight
Incident Management AWS Systems Manager Incident Manager

Azure

Component Azure Service / Tool
AI Risk Register Azure DevOps Boards; ServiceNow
Bias and Fairness Evaluation Azure Machine Learning Responsible AI Dashboard (Fairlearn)
Explainability Azure ML Interpretability (InterpretML)
Continuous Monitoring Azure ML Data Drift Monitor
Risk Dashboard Power BI + Azure Monitor
Incident Management Azure DevOps; PagerDuty

GCP

Component GCP Service / Tool
AI Risk Register Jira; Credo AI SaaS
Bias and Fairness Evaluation Vertex AI Explainable AI + What-If Tool
Continuous Monitoring Vertex AI Model Monitoring
Risk Dashboard Looker + Cloud Monitoring
Incident Management Cloud Ops + PagerDuty

On-Premises

Component Technology
AI Risk Register Archer; Resolver; custom PostgreSQL
Bias and Fairness Evaluation AI Fairness 360 (IBM open source); Fairlearn; Aequitas
Robustness Testing Adversarial Robustness Toolbox (IBM); TextAttack
Explainability SHAP; LIME; Alibi; Captum
Continuous Monitoring Evidently AI; WhyLabs; Grafana + Prometheus
Risk Dashboard Grafana; Metabase
Pattern ID Pattern Name Relationship Notes
EAAPL-CMP005 ISO 42001 Implementation COMPLEMENTARY ISO 42001 is the certifiable management system; NIST AI RMF provides detailed operational guidance; deploy together
EAAPL-CMP003 EU AI Act Compliance COMPLEMENTARY MAP function satisfies EU AI Act Article 9 risk assessment; MEASURE satisfies Article 61 monitoring
EAAPL-CMP002 APRA CPS234 AI Security COMPLEMENTARY GOVERN + MAP cover AI security risk; CPS234 provides additional security-specific obligations
EAAPL-CMP004 Privacy-Preserving AI COMPLEMENTARY MAP 3.1 impact on individuals includes privacy impacts; NIST AI RMF and Privacy Act both apply
EAAPL-AGT003 Human-in-the-Loop Oversight EXTENSION MANAGE 2.1 includes human oversight as a risk treatment; HITL pattern provides implementation
EAAPL-PLT007 AI Observability Platform PREREQUISITE MEASURE function requires operational AI monitoring infrastructure

17. Maturity Assessment

Overall Maturity Label: Proven

Dimension Level 1 Level 2 Level 3 Level 4 Level 5 Current Level
GOVERN No governance Informal governance AI Risk Committee; documented policy; RACI GOVERN actions 1.1–6.2 fully implemented Governance continuously improved; AI risk culture embedded Level 3
MAP No risk identification Ad-hoc risk identification Systematic MAP for all high-risk AI; Risk Profiles documented MAP integrated into AI project intake for all systems Real-time risk identification via automated context monitoring Level 3
MEASURE Performance metrics only Partial trustworthiness metrics All RMF trustworthiness characteristics measured Continuous automated measurement; risk scoring Predictive risk scoring; early warning system Level 3
MANAGE Reactive incident response Informal treatment plans Risk Treatment Plans for all risks above appetite; residual risk formally accepted Treatment effectiveness measured; feedback loop operational AI risk management continuously optimised via RMF profile Level 3

18. Revision History

Version Date Author Changes
1.0 2025-03-15 EAAPL Working Group Initial draft aligned to NIST AI RMF 1.0 (January 2023)
1.1 2025-09-01 EAAPL Working Group Added EO 14110 regulatory alignment; expanded MEASURE components
1.2 2026-06-12 EAAPL Working Group Added cascading failure scenario; updated reference implementations; aligned with NIST AI RMF Playbook v1.0 guidance
← Back to LibraryMore Regulatory Compliance →