7 Layers of AI Compliance: NIST AI RMF, ISO 42001 & SOC 2

AI Compliance Is Not One Framework

Most AI compliance programs fail because they pick one framework and stop. NIST AI RMF gets adopted as a policy document. ISO 42001 gets purchased as a cert. SOC 2 gets treated as an annual audit. None of them alone covers the full distance from legal obligation to production evidence.

The seven layers below are not a methodology I invented. They are the observable structure of what actually works in enterprises that pass audits and survive regulatory scrutiny. Each layer maps to a specific failure class that kills AI programs: legal blindness, framework without testing, testing without evidence, evidence without governance.

The NIST AI RMF provides the governance structure. ISO 42001 provides the management system. SOC 2 provides the controls audit. They are complementary, not competing. NIST tells you what to do. ISO tells you how to manage it. SOC 2 tells you whether your controls are working. (NIST AI RMF, ISO 42001, AICPA SOC 2)

For a deeper guide on operationalizing these frameworks, see how to secure and govern AI systems.

The Seven Layer Stack

Layer 7: Governance and Culture          <-- executive sponsorship
Layer 6: Evidence and Audit               <-- prove it on demand
Layer 5: Operations and Monitoring        <-- drift, incidents, improvement
Layer 4: Architecture and Infrastructure  <-- secure by design
Layer 3: Security Testing                 <-- find vulnerabilities first
Layer 2: Frameworks and Standards         <-- NIST, ISO, SOC 2
Layer 1: Legal and Regulatory             <-- the foundation

Direction of dependency: top depends on bottom
Direction of failure cascade: bottom breaks top

The layers are concurrent, not sequential. A new regulation in Layer 1 triggers changes in Layers 2 through 6. A security incident in Layer 3 requires evidence from Layer 6 and governance response from Layer 7. The stack forms a continuous loop, not a pipeline.

Layer 1: Legal and Regulatory

The foundation. This layer maps the laws and regulations that apply to your AI systems: EU AI Act, GDPR, HIPAA, TCPA, TRAIGA, NYC Local Law 144, and sector-specific rules. The output is an applicability assessment: which laws apply to which systems, and what each law requires.

As of 2026, the EU AI Act prohibits certain AI practices (Article 5) and imposes risk-based obligations on high-risk systems (Article 6-15). TRAIGA (Texas Responsible AI Governance Act) establishes an AI advisory council and prohibits certain governmental uses. NYC Local Law 144 requires bias audits for automated employment decision tools. (EU AI Act)

For regulatory-specific guides, see EU AI Act guide, Texas AI law guide, and NYC Local Law 144 guide.

Layer 2: Frameworks and Standards

Three frameworks form the core of AI compliance. They overlap but do not duplicate. Each addresses a different question.

NIST AI RMF 1.0 (published January 2023) provides four functions: Govern, Map, Measure, Manage. It tells you what to do across the AI lifecycle. (NIST AI RMF)

ISO/IEC 42001:2023 (published December 2023) is a management system standard for AI. It tells you how to operationalize AI governance through policies, procedures, and documented processes. (ISO 42001)

SOC 2 (AICPA Trust Services Criteria) is a controls audit. It tells you whether your controls are designed and operating effectively. SOC 2 is not AI-specific, but AI systems introduce new control considerations around data integrity, model behavior, and monitoring. (AICPA SOC 2)

Layer 3: Security Testing

Traditional security testing does not cover AI-specific attack surfaces. Prompt injection, RAG poisoning, model extraction, training data inference, and adversarial inputs require specialized testing approaches.

OWASP maintains the OWASP Top 10 for Large Language Model Applications, which catalogues the most critical LLM-specific vulnerabilities. MITRE ATLAS provides adversarial threat landscapes for AI systems. The Cloud Security Alliance AI Controls Matrix maps controls to AI-specific risks.

Organizations that adopt NIST AI RMF but never run adversarial tests have a framework without verification. The framework becomes documentation theater. For production RAG security patterns, see secure enterprise RAG architecture.

Layer 4: Architecture and Infrastructure

How your AI systems are built determines whether they are secure by design or secure by accident. This layer covers model selection, data pipelines, deployment infrastructure, integration patterns, and tenant isolation.

Architecture decisions made early are expensive to reverse. Multi-tenant isolation, RAG row-level security, model versioning, and audit logging are architectural concerns, not add-on features. For multi-tenant RAG isolation patterns, see implementing RAG row-level security for multi-tenant AI.

Layer 5: Operations and Monitoring

AI systems degrade over time. Models drift, data distributions shift, concepts evolve. Without monitoring, you are flying blind.

This layer covers drift detection, performance monitoring, incident response, and continuous improvement. The NIST AI RMF Measure function maps directly to this layer: you cannot manage what you do not measure. For logging pipeline architecture that satisfies EU AI Act Article 12, see building a logging pipeline for EU AI Act compliance.

Layer 6: Evidence and Audit

When an auditor asks you to prove compliance, this layer provides the answer. It is not enough to be compliant. You must be able to demonstrate it on demand.

Continuous evidence collection, audit trail generation, compliance reporting, and evidence retention all live here. For immutable audit trail implementation patterns, see implementing immutable audit trails for SOC 2 AI compliance.

Layer 7: Governance and Culture

AI governance committees, ethical review boards, stakeholder management, and organizational culture. This is the layer that makes everything else work. Without executive sponsorship and a culture of compliance, the other six layers become documentation exercises that nobody follows.

The NIST AI RMF Govern function and ISO 42001 Clause 5 (Leadership) both map to this layer. Governance is not a policy document. It is a decision-making structure with defined roles, escalation paths, and accountability.

Cross-Framework Control Mapping

The most common question I get from compliance officers is: how do NIST AI RMF, ISO 42001, and SOC 2 relate to each other? The mapping below shows how the three frameworks address the same underlying concerns from different angles.

This mapping is not exhaustive. It is a starting point for organizations that need to implement multiple frameworks without duplicating effort. The same evidence artifact (e.g., a drift detection log) can satisfy NIST Measure, ISO Clause 9.1, and SOC 2 CC4.1 simultaneously if it is structured correctly.

Common Gaps That Kill AI Programs

Three failure patterns appear repeatedly in enterprises that struggle with AI compliance:

Framework without testing. Organizations adopt NIST AI RMF as a policy but never run adversarial tests. The framework is documentation without verification. An auditor will ask for test results, not policy documents.

Testing without evidence. Security teams run penetration tests and adversarial evaluations but do not connect results to compliance reporting. The testing effort is wasted from an audit perspective because there is no traceable evidence chain.

Evidence without governance. Audit trails exist but nobody reviews them. Compliance becomes a checkbox exercise rather than a feedback loop. The governance layer (Layer 7) is what turns evidence into action.

Implementation Priority

Start with Layer 1: know which laws apply to your AI systems. Then work through the layers in parallel, not sequentially. The most expensive mistake is treating compliance as a linear pipeline where you finish one layer before starting the next.

If you need an independent systems advisor to assess your AI compliance program across all seven layers, schedule a strategic evaluation.

FAQ

What are the seven layers of AI compliance? The seven layers are: Legal and Regulatory, Frameworks and Standards (NIST AI RMF, ISO 42001, SOC 2), Security Testing, Architecture and Infrastructure, Operations and Monitoring, Evidence and Audit, and Governance and Culture. Each layer addresses a specific failure class in AI compliance programs.

How do NIST AI RMF, ISO 42001, and SOC 2 relate? NIST AI RMF tells you what to do across the AI lifecycle (Govern, Map, Measure, Manage). ISO 42001 tells you how to manage it through a documented management system. SOC 2 tells you whether your controls are designed and operating effectively. They are complementary, not competing.

Is NIST AI RMF mandatory? NIST AI RMF is a voluntary framework, not a regulation. However, it is increasingly referenced by regulators and procurement requirements. Some federal agencies and contracts require alignment with NIST AI RMF. Check your specific regulatory and contractual obligations.

Does SOC 2 cover AI systems? SOC 2 is not AI-specific, but AI systems introduce new control considerations around data integrity, model behavior, monitoring, and incident response. Organizations deploying AI need to extend their SOC 2 controls to address these AI-specific risks.

What is the most common AI compliance gap? The most common gap is framework without testing. Organizations adopt NIST AI RMF as a policy document but never run adversarial tests or security evaluations. The framework becomes documentation theater without verification evidence.

How do I map NIST AI RMF to ISO 42001? NIST AI RMF Govern maps to ISO 42001 Clause 5 (Leadership) and Clause 6 (Planning). Map maps to Clause 6.1 (Risk actions). Measure maps to Clause 9 (Performance evaluation). Manage maps to Clause 10 (Improvement). Use the cross-framework mapping table in this article as a starting point.