Regulatory Policy Mapping for AI Agents: From EU AI Act to NIST AI RMF to Production Rules

Regulatory compliance for AI systems is transitioning from aspiration to requirement. The EU AI Act reached full applicability for high-risk AI systems in August 2026. NIST AI RMF 1.0 has become the de facto reference framework for AI governance in the United States, with increasing adoption as a vendor qualification criterion in federal procurement. ISO 42001 provides the AI management system standard that ISO 27001-certified organizations are extending to cover their AI deployments.

The challenge for organizations deploying AI agents is that these frameworks are written in the language of governance — risk management systems, accountability structures, documentation requirements — while operational teams need to translate them into the language of engineering: specific policies, enforcement mechanisms, audit logs, and testable controls.

This document provides the definitive translation guide. We map each major regulatory framework to concrete operational agent policies, provide gap analysis methodology for identifying where an organization's current posture falls short, and describe the compliance automation patterns that make ongoing regulatory adherence achievable without a full-time compliance team for every agent deployment.

---

TL;DR

• The EU AI Act, NIST AI RMF, and ISO 42001 share common themes but use different vocabulary and have different scopes — understanding the differences is prerequisite to mapping them to operational policies.
• EU AI Act compliance for AI agents depends on risk tier: most enterprise AI agents qualify as "limited risk" or "general purpose AI" under the Act; high-risk AI system requirements are more stringent.
• NIST AI RMF's four functions (Govern, Map, Measure, Manage) map to specific policy categories: Govern → policy governance, Map → risk-scoped policies, Measure → monitoring policies, Manage → incident response policies.
• ISO 42001 provides the management system structure; individual AI agent policies are the controls that satisfy ISO 42001 requirements.
• Gap analysis requires mapping each regulatory requirement to either an existing policy (satisfied), a partially-addressed gap (requires enhancement), or an open gap (requires new policy).
• Compliance automation — machine-readable regulatory requirements, automated policy verification, continuous evidence collection — converts point-in-time compliance into continuous compliance.
• Armalo's trust oracle provides third-party attestation of AI agent behavioral properties that supports regulatory compliance evidence requirements.

---

Regulatory Landscape Overview

EU AI Act: Risk-Tiered Requirements

The EU AI Act classifies AI systems into four risk tiers, each with different compliance requirements:

Unacceptable risk (prohibited): AI systems that pose unacceptable risks to fundamental rights. Examples: social scoring systems by public authorities, real-time biometric surveillance in public spaces. No enterprise AI agents in standard deployments fall in this category.

High risk: AI systems in domains specified in Annexes II and III, including: critical infrastructure management, educational assessment, employment and worker management, access to essential public services, law enforcement, migration management, justice administration. High-risk requirements include: risk management systems, data governance, technical documentation, transparency, human oversight, accuracy/robustness/cybersecurity, and registration in the EU database.

Limited risk: AI systems that interact with humans or generate content, with transparency obligations. An AI agent that interacts with customers must disclose that it is an AI system.

Minimal risk: All other AI systems, for which compliance is voluntary best practice. Most internal enterprise AI agents fall here.

General Purpose AI Models (GPAI): Foundation models above capability thresholds have additional requirements including technical documentation, copyright compliance, and transparency.

NIST AI RMF: A Framework, Not Regulations

NIST AI RMF 1.0 is a voluntary framework, not a regulation. Its importance lies in its adoption as a compliance reference:

• U.S. federal agencies increasingly require NIST AI RMF alignment in AI system procurements
• NIST AI RMF is referenced in the EU AI Act's Article 40 as a relevant standard
• Many enterprises are using NIST AI RMF as the internal governance standard for their AI programs

The four core functions:

• Govern: Policies, processes, organizational roles, and culture for AI risk management
• Map: Identify and prioritize AI risks in context
• Measure: Assess, analyze, and track AI risk
• Manage: Prioritize, respond to, and communicate AI risks

ISO 42001: The Management System Standard

ISO 42001:2023 is the international standard for AI management systems. It follows the Annex SL structure used by ISO 27001, ISO 9001, and other management system standards — which means organizations already certified under other ISO standards have a familiar framework for ISO 42001 implementation.

ISO 42001 requires:

• Context establishment (internal and external context for AI systems)
• Leadership commitment and AI policy
• Planning (risk assessment and treatment, AI objectives)
• Support (resources, competence, awareness, communication, documented information)
• Operations (operational planning and control, AI impact assessment)
• Performance evaluation (monitoring, measurement, analysis, evaluation, internal audit, management review)
• Improvement (nonconformity, continual improvement)

---

EU AI Act: Article-by-Article Mapping for AI Agents

Article 9: Risk Management System

Requirement: High-risk AI systems must implement a continuous risk management system throughout the AI system lifecycle, including risk identification, estimation, evaluation, and risk control measures.

Policy mapping:

Article 9 → Risk management policies:
- Policy: mandatory_risk_assessment_before_deployment
  Trigger: new agent deployment or significant capability change
  Action: require risk assessment document, approve before deploy
  
- Policy: continuous_risk_monitoring
  Trigger: continuous
  Action: monitor behavioral metrics vs. risk thresholds; alert on threshold breach
  
- Policy: residual_risk_acceptance
  Trigger: after risk controls applied
  Action: require documented acceptance of residual risks above threshold

Evidence requirements: Risk assessment documents, risk control implementation records, periodic risk review records.

Article 10: Data and Data Governance

Requirement: Training, validation, and testing datasets must meet quality criteria; data governance and management practices must be in place; datasets must be relevant, representative, and free of errors.

Policy mapping:

Article 10 → Data governance policies:
- Policy: training_data_provenance_required
  Scope: all models used in high-risk AI deployments
  Requirement: documented provenance for all training datasets
  
- Policy: dataset_quality_gate
  Trigger: model onboarding
  Action: require dataset quality assessment; block deployment if assessment missing
  
- Policy: data_bias_assessment
  Trigger: model onboarding; periodic reassessment
  Action: require bias testing documentation across defined demographic categories

Article 13: Transparency and Information Provision

Requirement: High-risk AI systems must be transparent enough to enable deployers to interpret outputs and use the system appropriately.

Policy mapping:

Article 13 → Transparency policies:
- Policy: explainability_for_high_stakes_decisions
  Trigger: agent decision above consequence threshold
  Action: require human-readable explanation to accompany decision
  
- Policy: decision_uncertainty_disclosure
  Trigger: agent confidence below threshold
  Action: require disclosure of low-confidence state to end user
  
- Policy: ai_identity_disclosure
  Trigger: agent interaction with natural persons
  Action: require disclosure that the interacting system is AI

Article 14: Human Oversight

Requirement: High-risk AI systems must be designed with effective human oversight during the period of use, with tools enabling overseers to understand the AI system, pause operation, and override outputs.

Policy mapping:

Article 14 → Human oversight policies:
- Policy: oversight_gate_high_consequence
  Trigger: agent action above consequence tier 3
  Action: require human approval before execution
  
- Policy: override_capability_required
  Scope: all agents with action capability
  Requirement: operator must be able to pause agent operations within 60 seconds
  
- Policy: audit_log_human_readable
  Scope: all agent decisions in oversight scope
  Requirement: audit log format must be interpretable by oversight role without
               specialist training

Article 15: Accuracy, Robustness, and Cybersecurity

Requirement: High-risk AI systems must achieve appropriate levels of accuracy, robustness, and cybersecurity throughout their lifecycle.

Policy mapping:

Article 15 → Security and robustness policies:
- Policy: accuracy_baseline_required
  Trigger: model deployment
  Action: require documented accuracy benchmarks on representative test set
  
- Policy: adversarial_robustness_testing
  Trigger: deployment; periodic (minimum semi-annual)
  Action: require red team adversarial testing; document results
  
- Policy: cybersecurity_controls_required
  Scope: all high-risk AI deployments
  Requirement: satisfy organizational security baseline controls
               (injection defense, tool permission hardening, egress controls)

---

NIST AI RMF: Function-to-Policy Mapping

Govern Function

NIST AI RMF Govern covers organizational policies, roles, and culture. Corresponding policy categories:

Govern 1.1 — Policies for AI risk management:

→ Policy: ai_policy_governance_structure
  Requirement: documented AI policy ownership, review cycle, and enforcement authority
  
→ Policy: ai_risk_appetite_statement
  Requirement: documented organizational risk tolerance levels for AI systems

Govern 2.2 — Roles and responsibilities:

→ Policy: role_accountability_assignment
  Requirement: for each AI agent deployment, documented owner, developer, deployer,
               and oversight roles
               
→ Policy: ai_incident_escalation_path
  Requirement: documented escalation path from agent-level incident to CISO/CTO level

Govern 5.1 — Ongoing monitoring:

→ Policy: continuous_behavioral_monitoring
  Requirement: automated monitoring with documented metrics, thresholds, and response procedures
  
→ Policy: periodic_alignment_review
  Trigger: minimum quarterly
  Action: review agent behavioral metrics against pact commitments; document findings

Map Function

NIST AI RMF Map covers context establishment and risk identification. Corresponding policy categories:

Map 1.6 — Organizational risk policies:

→ Policy: ai_system_classification
  Trigger: new AI deployment
  Action: classify system by NIST risk category; assign appropriate policy tier
  
→ Policy: stakeholder_impact_assessment
  Trigger: new AI deployment or significant change
  Action: document affected stakeholders and potential impacts

Map 5.1 — Likelihood and impact assessment:

→ Policy: pre_deployment_risk_register
  Trigger: new AI deployment
  Action: complete risk register including likelihood and impact ratings;
          require sign-off before production

Measure Function

NIST AI RMF Measure covers risk assessment, analysis, and tracking. Corresponding policy categories:

Measure 2.1 — Test and evaluation:

→ Policy: pre_deployment_behavioral_testing
  Trigger: new deployment; major changes
  Action: complete behavioral test suite; document results; gate on pass criteria
  
→ Policy: adversarial_evaluation
  Trigger: deployment; minimum semi-annual
  Action: adversarial red team exercise; document findings; require remediation plan

Measure 4.1 — Performance metrics:

→ Policy: behavioral_metrics_monitoring
  Requirement: documented set of behavioral metrics, baselines, and alert thresholds;
               active monitoring in production

Manage Function

NIST AI RMF Manage covers risk response and communication. Corresponding policy categories:

Manage 1.1 — Risk response plans:

→ Policy: ai_incident_response_plan
  Requirement: documented incident response procedures for AI agent failures,
               including quarantine, investigation, and remediation steps

Manage 4.1 — Post-incident review:

→ Policy: post_incident_review_required
  Trigger: P1 or P2 AI agent incident
  Action: require post-incident review within 5 business days;
          document root cause; update policies as needed

---

ISO 42001: Clause-by-Clause Policy Requirements

Clause 4: Context of the Organization

4.1 — Understanding the organization and its context:

→ Policy: ai_context_documentation
  Requirement: document internal and external factors affecting AI risk
  Review cycle: annual or when significant changes occur

4.2 — Understanding needs and expectations of interested parties:

→ Policy: stakeholder_requirements_documentation
  Requirement: document regulatory requirements, customer requirements, and
               organizational requirements applicable to AI systems

Clause 6: Planning

6.1.2 — AI risk assessment:

→ Policy: ai_risk_assessment_process
  Requirement: documented risk assessment methodology; applied before deployment
               and at defined intervals

6.1.3 — AI risk treatment:

→ Policy: risk_treatment_plan
  Requirement: for each identified risk above threshold, documented treatment plan
               with owner, deadline, and success criteria

Clause 8: Operations

8.4 — AI system impact assessment:

→ Policy: impact_assessment_required
  Trigger: new deployment; changes to scope or capabilities
  Action: complete AI impact assessment covering: affected parties, potential harms,
          severity/likelihood, mitigations

Clause 9: Performance Evaluation

9.1 — Monitoring and measurement:

→ Policy: iso42001_monitoring_metrics
  Requirement: documented monitoring metrics, measurement methods, and
               evaluation frequency for all ISO 42001 objectives

---

Gap Analysis Methodology

Step 1: Requirements Inventory

Create a comprehensive inventory of applicable regulatory requirements. For each applicable framework:

• List every applicable article, section, or control
• Note the requirement type: policy, process, documentation, technical control
• Note the evidence required to demonstrate compliance

Step 2: Current State Assessment

For each requirement, assess the current state:

• Satisfied: An existing policy, process, or technical control clearly satisfies the requirement; evidence is available.
• Partial: The requirement is partially addressed; gaps remain; evidence is incomplete.
• Gap: No policy, process, or control exists for this requirement.

Step 3: Gap Prioritization

Prioritize gaps by:

1. Regulatory enforcement risk (gaps that, if discovered by a regulator, carry significant penalties)
2. Security risk (gaps that, if exploited, cause significant harm)
3. Implementation complexity (high-impact, low-complexity gaps first)

Step 4: Remediation Planning

For each gap:

• Define the policy or control required to close the gap
• Assign ownership
• Define the implementation timeline
• Define the evidence that will demonstrate the gap is closed

Step 5: Evidence Collection Automation

Map each regulatory requirement to the automated evidence that the policy management system can produce:

• Audit logs demonstrating policy enforcement events
• Test results demonstrating policy effectiveness
• Monitoring reports demonstrating ongoing compliance
• Review records demonstrating governance activities

Automate evidence collection wherever possible — manually collected evidence is error-prone and unsustainable at scale.

---

Compliance Automation Architecture

Manual compliance processes fail at scale. A policy management system designed for compliance automation provides:

Machine-readable regulatory requirements: Regulatory requirements encoded in structured format (not prose) that can be evaluated against policy artifacts automatically.

Policy-to-requirement traceability: Every policy in the policy management system links to the regulatory requirements it satisfies. A change to a regulatory requirement triggers automatic identification of all affected policies.

Automated evidence collection: Every policy enforcement event is logged with sufficient metadata to serve as compliance evidence. Reports can be generated on demand.

Continuous compliance monitoring: Dashboard showing current compliance status across all applicable regulatory requirements, with alerts when compliance posture changes.

Audit package generation: At audit time, automatically generate a package of evidence artifacts demonstrating compliance with each applicable requirement.

---

How Armalo Supports Regulatory Compliance Evidence

Armalo's trust oracle provides the third-party behavioral attestation that many regulatory frameworks require as evidence of effective AI governance. When EU AI Act Article 15 requires evidence of cybersecurity testing, or when NIST AI RMF Measure 2.1 requires evidence of test and evaluation, Armalo's evaluation records — showing the adversarial testing conducted, the techniques tested, and the agent's behavioral outcomes — provide independently verified evidence that is more compelling than self-attestation.

The behavioral pact mechanism creates a documented, version-controlled record of the agent's declared behavioral commitments — satisfying documentation requirements under Article 11 (Technical Documentation) and ISO 42001 Clause 7.5 (Documented Information).

The composite trust score provides the quantified behavioral baseline required by NIST AI RMF Measure 2.8 and Measure 4.1 — a documented measurement of AI system performance across defined dimensions, tracked over time.

---

Conclusion: Regulatory Compliance as Engineering

The organizations that will handle the coming wave of AI regulation with the least disruption are those that treat regulatory compliance as an engineering problem — one that can be decomposed into specific requirements, mapped to specific controls, automated with specific tools, and verified with specific metrics.

The frameworks described here — EU AI Act, NIST AI RMF, ISO 42001 — are rigorous, but they are not arbitrary. They represent the consensus of regulators, standards bodies, and AI governance practitioners about what makes AI systems trustworthy enough to deploy in consequential contexts. An organization that builds the policy infrastructure to satisfy these requirements will have built AI agent governance infrastructure that genuinely reduces risk — not merely satisfies compliance checklists.

The translation from regulatory language to operational policy is the work described in this document. The organizations that do this translation rigorously, and maintain it as regulations evolve, will be the ones that can deploy AI agents in regulated industries with confidence — and that can demonstrate that confidence to regulators, customers, and the public.