AI Compliance: ISO 42001, SOC 2, HIPAA, ISO 27001 & EU AI Act Explained

Why AI Compliance Is Harder Than Traditional IT Compliance

Traditional compliance frameworks were designed for deterministic systems — software that does the same thing every time given the same inputs. Auditors could review code, test procedures, and sample logs to verify that controls were operating as designed.

AI agents break that model. They are non-deterministic, their reasoning is often opaque, and their action scope can span many systems simultaneously. Compliance teams trying to apply traditional frameworks to AI systems quickly discover gaps: existing controls don't map cleanly, evidence collection is unclear, and auditors are asking questions the organization doesn't yet have answers for.

The good news: the core principles of every major compliance framework — access control, audit trails, risk management, incident response — apply directly to AI agents. What's needed is a translation layer that maps those principles to AI-specific controls.

ISO 42001:2023 — The AI Management System Standard

ISO 42001:2023 is the world's first international standard specifically for AI Management Systems (AIMS). Published in December 2023 by the International Organization for Standardization, it provides organizations with a structured framework for responsible development, deployment, and governance of AI systems — including autonomous AI agents.

Unlike ISO 27001 (information security) or SOC 2 (service organization controls), ISO 42001 was designed from the ground up for AI. It addresses the unique challenges of AI governance: algorithmic transparency, AI-specific risk assessment, human oversight of automated decisions, and the lifecycle management of AI systems.

What ISO 42001 requires for AI agent deployments:

• AI policy and objectives (Clause 5): Organizations must establish a formal AI policy defining how AI will be used, what values guide AI decisions, and how AI governance aligns with business objectives. For AI agents, this means documented policies on autonomy limits, human oversight thresholds, and acceptable use cases.
• AI risk assessment (Clause 6): A systematic process for identifying and evaluating AI-specific risks before deployment. This goes beyond information security risk — it includes risks of biased outputs, unintended agent behaviors, and misuse. Every AI agent deployment should have a documented risk assessment on file.
• Operational controls (Clause 8): Processes to ensure AI systems behave as intended throughout their lifecycle. For agents, this includes system prompt management, tool authorization controls, behavioral monitoring, and version control for model configurations.
• Impact assessment (Annex B): ISO 42001 includes guidance on assessing the societal and individual impacts of AI systems — particularly relevant for AI agents making decisions that affect employees, customers, or third parties.
• Transparency and explainability: Documentation that allows affected parties to understand how AI decisions are made. For enterprise AI agents, this requires immutable execution traces and the ability to reconstruct agent reasoning for any given decision.
• Human oversight mechanisms: Controls that keep humans informed and in control of consequential AI actions. Approval queues, escalation paths, and override capabilities are all required for high-impact agent operations.

Certification path: ISO 42001 supports third-party certification, similar to ISO 27001. Organizations can pursue certification through an accredited certification body. Many enterprises are seeking ISO 42001 certification alongside ISO 27001 to demonstrate comprehensive governance of both information security and AI systems.

Key evidence to collect: AI policy documentation, risk assessment records for each agent, operational control procedures, impact assessments for high-risk deployments, human oversight logs showing escalation decisions, and training records for AI governance personnel.

SOC 2 and AI Agents

SOC 2 audits evaluate an organization's controls around security, availability, processing integrity, confidentiality, and privacy. For AI agents, the most relevant criteria are Security (CC6, CC7, CC8) and Confidentiality (C1).

What SOC 2 auditors look for with AI systems:

• Logical access controls (CC6.1): Evidence that AI agents only have access to systems and data authorized for their specific function. Access reviews must be documented.
• Change management (CC8.1): Evidence that new AI agents go through a review process before deployment, with documented approval.
• Anomaly detection (CC7.2): Monitoring systems that detect unusual AI agent behavior and defined response procedures when anomalies are identified.
• Incident response (CC7.3–7.5): Documented procedures for AI-related security events with evidence that procedures were followed when incidents occurred.
• Vendor management (CC9.2): If AI agents use third-party model providers (OpenAI, Anthropic), those vendors must be assessed and monitored.

Key evidence to collect: Agent access logs, deployment approval records, anomaly alert history and response documentation, vendor risk assessments for AI providers.

HIPAA and AI Agents in Healthcare

HIPAA's Security Rule requires administrative, physical, and technical safeguards for systems that create, receive, maintain, or transmit electronic protected health information (ePHI). Any AI agent that touches patient data falls under these requirements.

HIPAA requirements most relevant to AI agents:

• Access control (§164.312(a)(1)): Each AI agent must have a unique identifier and access only the ePHI required for its specific function.
• Audit controls (§164.312(b)): Mechanisms to record and examine activity on AI systems that handle ePHI. This requires full execution traces, not just error logs.
• Transmission security (§164.312(e)(1)): Any ePHI transmitted to external AI providers (including model inference calls) must be encrypted and covered by a Business Associate Agreement (BAA). Most major AI providers offer BAAs — verify before using any AI service with patient data.
• Minimum necessary standard: AI agents must only access the minimum ePHI needed for the specific task. Broad database access permissions are non-compliant even if the agent only queries a small subset.

Critical point: Sending ePHI in a prompt to a cloud AI provider without a BAA is a HIPAA violation — regardless of whether the provider encrypts the data in transit. Always verify BAA status before connecting any patient data source to an AI agent.

ISO 27001:2022 and AI

The 2022 revision of ISO 27001 added Annex A controls that explicitly address AI and cloud environments. Organizations seeking or maintaining ISO 27001 certification now need to demonstrate controls specifically for AI systems.

Relevant Annex A controls for AI agents:

• A.5.23 — Information security for use of cloud services: Policies for AI agents using cloud-based model providers, including data handling, access control, and exit procedures.
• A.8.9 — Configuration management: Documented, controlled configurations for AI agents including model versions, system prompts, and tool permissions.
• A.8.16 — Monitoring activities: Monitoring of AI agent behavior with defined baselines for normal operation and alerting thresholds for deviations.
• A.8.25 — Secure development lifecycle: Security reviews for AI agents as part of their development and deployment process, equivalent to secure code review for traditional software.

EU AI Act

The EU AI Act is the world's first comprehensive AI regulation. It categorizes AI systems by risk level and applies progressively more stringent requirements to higher-risk systems.

Risk categories relevant to enterprise AI agents:

• High-risk AI (Article 6): AI systems used in employment decisions, credit scoring, essential services, law enforcement, or critical infrastructure. These require conformity assessments, technical documentation, human oversight mechanisms, accuracy and robustness requirements, and registration in the EU AI database.
• Limited risk: AI systems that interact with humans must disclose they are AI (transparency requirement). Most customer-facing AI agents fall here at minimum.
• General purpose AI (GPAI) models: Foundation model providers (OpenAI, Anthropic, Google) have their own obligations. Enterprise deployers using these models as the basis for high-risk applications inherit obligations to implement appropriate controls.

Enforcement timeline: High-risk AI system obligations under the EU AI Act are now enforceable. Organizations deploying AI agents in the EU — or making AI-based decisions about EU residents — should assess compliance now.

The Controls That Satisfy All Five Frameworks Simultaneously

Rather than building separate compliance programs for each framework, most enterprises can satisfy ISO 42001, SOC 2, HIPAA, ISO 27001, and the EU AI Act with a common set of AI governance controls:

• Complete agent inventory with ownership, data access scope, and deployment approval records
• Least-privilege access controls reviewed quarterly and documented
• Immutable execution traces with minimum 1-year retention
• PII/PHI detection and redaction on all agent inputs and outputs
• Policy enforcement logs showing every policy evaluation and outcome
• Incident response documentation with AI-specific runbooks
• Vendor risk assessments for all AI model providers including BAA status
• Human oversight mechanisms for high-risk AI decisions