And Why It Requires an Information Model, Not Just a Policy Framework
The Problem
Organizations use the phrase "AI governance" to describe many different activities: intake questionnaires, policy statements, security reviews, model evaluations, privacy assessments, and approval workflows. All of these may be part of AI governance, but none of them, standing alone, defines it. Without a shared definition, governance becomes a label for disconnected review steps rather than a coherent operating model.
The major governance frameworks - NIST AI RMF, ISO/IEC 42001, the EU AI Act - describe what organizations should accomplish: identify risks, assess systems, produce accountable decisions. But they do not define the information model needed to make those activities rigorous. They tell you to map your systems without specifying what a map looks like. As a result, most organizations fill the gap with questionnaires and stakeholder attestations, which capture what people believe about a system rather than what the system actually does.
A Working Definition
AI governance is the cross-functional operating model used to describe AI systems as deployed, evaluate them against an organization's policy and risk priorities, and produce accountable decisions about approval, mitigation, escalation, and residual risk.
AI systems require this kind of governance because their risks do not arise from a single model or component in isolation. They arise from the combination of non-deterministic model behavior and the integrated behavior of the deployed system: what data the system receives, what tools it can invoke, what resources it can reach, what constraints apply along the way, and what downstream actions its outputs can trigger.
AI governance is therefore not a queue, a questionnaire, or a set of policy statements. It is the process by which an organization turns a real AI deployment into a shared object of review and then into a reasoned governance outcome.
Two Layers: Description and Evaluation
Effective governance requires two distinct layers. Getting this structure right is the difference between governance that produces defensible decisions and governance that produces paperwork.
The Information Layer: Describing the System
Before an organization can evaluate an AI system, it needs a structured, factual description of that system. This is the same principle that makes other governance domains work: privacy governance depends on data mapping; security governance depends on asset inventories and threat models. AI governance needs an equivalent.
The CORE Framework provides that information model. It describes an AI system in terms of four elements:
- Components - the models, tools, APIs, databases, and other functional units that make up the system.
- Operations - what those components can do, including the actions, transformations, and decisions they perform.
- Resources - the data, credentials, external services, and other assets the system can access.
- Execution - how control and information flow through the system, including orchestration patterns, handoffs, constraints, and feedback loops.
In practice, CORE creates a common system record - a blueprint - that captures the relevant components, the relationships among them, the paths through which information and actions move, and the controls that exist along those paths. This is the minimum factual foundation needed for meaningful governance.
The Evaluation Layer: Policies, Risks, and Outcomes
Once a system description exists, evaluation proceeds through three activities:
Policies. Clear statements of what the organization expects to be true of its AI systems. Some are inherited from privacy, security, and compliance practice. Others are AI-specific: restrictions on agentic behavior, requirements for human oversight in particular contexts, guardrail expectations for external actions. Where possible, policies are expressed as code and evaluated automatically against system blueprints.
Risks. A structured taxonomy of the risks that apply to AI systems. These include conventional security threats, privacy and data governance risks, and AI-specific failure modes such as hallucination, prompt injection, uncontrolled tool use, or emergent behavior in multi-agent systems. A comprehensive risk registry classifies risks across categories and maps them to applicable frameworks, enabling organizations to select the risks relevant to their priorities and regulatory obligations.
Outcomes. The purpose of governance is not to inventory systems or label risks. It is to produce a decision and an evidentiary record. Outcome analysis uses an FMEA-inspired methodology (Severity x Occurrence x Detection) to evaluate each applicable risk in the context of the actual deployed system, producing Risk Priority Numbers that ground mitigation decisions in structured reasoning rather than subjective judgment.
Alignment with Major Frameworks
This approach is designed to operationalize, not replace, established governance frameworks. The table below illustrates how it maps to NIST AI RMF. The same structure supports ISO/IEC 42001 compliance and EU AI Act conformity assessment.
| NIST AI RMF | CORE + Governance Layer | What It Produces |
|---|---|---|
| GOVERN | Organizational risk priorities; policy library | Risk scope and evaluation criteria |
| MAP | CORE system description (Components, Operations, Resources, Execution) | Structured, reviewable system blueprint |
| MEASURE | Policy evaluation against blueprints (automated where possible) | Findings, control gaps, risk scores |
| MANAGE | FMEA-based outcome analysis; RPN scoring | Decisions, mitigations, evidentiary record |
Integration, Not Replacement
AI governance does not replace privacy governance, security governance, or compliance functions. It gives them a common system-level object of review. Privacy governance remains authoritative for privacy requirements and privacy risk judgments. Security governance remains authoritative for security requirements and security risk judgments. But each operates more effectively when working from a shared, structured description of the system under review.
Existing intake and workflow tools remain important for routing, accountability, and recordkeeping. But they are not, by themselves, AI governance. AI governance is the operating model that allows privacy, security, legal, compliance, and business stakeholders to assess the same deployed system in a coordinated way and reach a coherent result.
Why This Matters Now
Regulatory requirements are converging on a common expectation: that organizations can demonstrate - not merely assert - that their AI systems are governed. The EU AI Act requires conformity assessment for high-risk systems. ISO 42001 requires documented risk treatment and lifecycle management. NIST AI RMF calls for continuous mapping and measurement. State-level legislation in the U.S. is adding disclosure, assessment, and accountability obligations.
All of these requirements assume the organization knows what its AI systems do. Without a structured information model, that assumption fails. Organizations that build governance on questionnaires and attestations will struggle to demonstrate compliance in an environment where regulators increasingly audit the backend, not the banner.