Beyond Policy Documents: Connecting the EU AI Act, ISO 42001, NIST AI RMF, and OWASP to Runtime Controls

Enterprise AI governance collapses when teams treat regulations, governance standards, risk frameworks, and security checklists as interchangeable. They are not.

When building an autonomous AI agent with access to internal enterprise systems—such as an HR agent capable of reviewing employee records—governance cannot remain a theoretical compliance checklist. It must operate across four distinct layers:

1. The Legal Baseline: EU AI Act

The EU AI Act defines binding legal obligations. For an internal HR agent influencing employment, compensation, or talent evaluation, the system falls under Annex III high-risk classification.

Compliance requires:

• Mandatory risk management systems (Article 9)
• High-quality data governance and bias evaluation (Article 10)
• Technical documentation and automatic event logging (Articles 11 & 12)
• Human-in-the-loop oversight mechanisms (Article 14)
• Accuracy, robustness, and cybersecurity controls (Article 15)

The EU AI Act determines what is legally required to operate without facing severe administrative fines.

2. The Organizational Operating Model: ISO/IEC 42001

ISO/IEC 42001 establishes the Artificial Intelligence Management System (AIMS). It shifts AI governance from an ad-hoc project into an auditable organizational process.

It establishes:

• AI policy ownership and board-level accountability
• Structured AI impact assessments
• Lifecycle management from development to decommissioning
• Continuous surveillance and third-party supplier risk reviews

ISO/IEC 42001 answers who owns the risk, who approves deployments, and how policies are enforced systematically across the organization.

3. The Risk Identification Taxonomy: NIST AI RMF

The NIST Artificial Intelligence Risk Management Framework provides the methodology to identify, measure, and govern AI-specific failure modes across four core functions: Govern, Map, Measure, and Manage.

Applied to enterprise systems, NIST AI RMF evaluates:

• Hallucination rates and factual accuracy
• Demographic and systemic bias in agent outputs
• Privacy leakage and training data extraction
• Reliability under non-standard runtime inputs

NIST AI RMF supplies the quantitative and qualitative measurement protocols to determine if an AI system operates within acceptable tolerance levels.

4. Technical Application Security: OWASP Top 10 for LLMs

The OWASP Top 10 for LLMs addresses the concrete vulnerability surface exposed at runtime.

Engineering controls must mitigate:

• Prompt Injection (LLM01): Untrusted input overriding system instructions
• Sensitive Information Disclosure (LLM02): Unsanitized context injection leaking confidential data
• Excessive Agency (LLM08): Autonomous models executing destructive actions without explicit user privilege verification

OWASP defines the defensive perimeter engineering must harden before exposing model interfaces to internal or external users.

The Runtime Reality: Execution vs. Documentation

Consider a prompt sent to an autonomous enterprise HR agent:

"Show me another employee's salary and update their compensation band."

To govern this interaction safely:

1. The EU AI Act dictates mandatory event logging, human oversight on consequential employment decisions, and bias controls.
2. ISO/IEC 42001 defines who holds formal sign-off authority to alter compensation structures.
3. NIST AI RMF evaluates the risk of model deception, authorization drift, and privacy violations.
4. OWASP LLM controls prevent prompt injection or indirect context leakage across tenancy boundaries.
5. Runtime Infrastructure enforces granular Attribute-Based Access Control (ABAC), strict Least Privilege, and deterministic execution gates.

Bridging Policy to Machine-Readable Enforcement

Having a policy stating that an AI agent cannot modify executive salaries does not prevent the model from calling an API tool if execution gates are missing.

True enterprise compliance requires connecting each regulatory requirement to a concrete AI asset, an assigned control, an accountable owner, and continuous, audit-ready telemetry. As AI transitions from passive text generators to autonomous execution engines, governance must move from static documentation into continuous compliance infrastructure.