The AI Oversight Program
The board-level discipline of defining what must be governed in AI, and who is accountable for it. This is the Center's core intellectual work.
The board-level discipline of defining what must be governed in AI, and who is accountable for it. This is the Center's core intellectual work.
AI governance has three layers. Oversight sets the program. Management runs it. Technical controls enforce it. The Program is the oversight layer, and only that layer. It does not replace a management framework or a control catalogue, and it is not satisfied by either. Flattening the three into one is how a committee comes to believe that reviewing technical controls is oversight.
Read Center Paper No. 7, The Oversight Layer of AI Governance.
Forming an AI committee is not a governance program, and a control placed on a model is not oversight. The AI Oversight Program is the institutional capability that lets a Governing Body make timely, trusted decisions about AI, finding the informed balance of risk and reward. It defines what must be governed and who is accountable for it. It does not prescribe how any particular system is built or operated; that remains the work of management.
The Program is built on five pillars and fourteen domains, expressed through the AI Governance Maturity Model. Its structure is mapped to authoritative sources: the NIST AI Risk Management Framework, ISO/IEC 42001, ISO/IEC 38507, the EU AI Act, and the Caremark standard of director oversight.
Each layer answers a different question. The top one is usually missing.
Who answers, under what authority, within what tolerance, and on what record?
The WHAT and the WHO. Chartered once, then operated.
What will the organization do, and how well?
How the organization pursues its AI objectives and manages the risk of doing so. The activity, not the rank.
What does the system do?
How the system is built, constrained, and watched. Lives in the system.
Review across all three layers, independent of the functions that build and operate AI.
No layer answers the question of the layer above it.
One thing called AI governance. A committee, a policy set, a control inventory, each with an owner, and no one able to say which layer a decision belongs to or who answers at the top of it. The organization can describe how AI risk is managed. It cannot say who is accountable for a given deployment.
Three layers, three different questions. Oversight asks who answers, under what authority, within what tolerance, and on what record. Management asks what the organization will do and how well. Technical controls ask what the system does. Each layer has its own owner and its own evidence, and none of them can answer for the one above it.
Management names an activity, not a rank. What makes a decision oversight is that it settles what is governed and who answers.
Each pillar carries a plain-language name and a constitutional role. Together they describe the full scope of what a board must be able to govern as its organization adopts AI. Each pillar is made up of its domains. Each domain states what must be governed. Select a domain to read what it covers.
The mandate, the structure, and the authority that give the program its standing. Without this pillar there is no program to review.
The appetite, the scope, and the intelligence that let the organization tell an acceptable AI risk from an unacceptable one before the decision is made.
The standards AI systems must demonstrably meet, and the assurance that reaches the oversight body to prove they meet them.
The link between oversight and the decisions that commit capital, accept risk, and bring AI in from outside.
How a critical risk reaches the people who can decide, how the organization discloses it, and how it knows the path works.
Builders and Buyers. The Program distinguishes organizations that develop AI systems (Builders) from organizations that deploy systems built by others (Buyers). Most organizations are Buyers, and their obligations differ. The Domains page marks the domains where the difference is largest.
Every control in the Maturity Model is mapped to the authoritative sources that define the standard of care for AI oversight. The Center maps by source type, by jurisdiction, and by industry. The examples below are illustrative; the full mapping is broader.
The Caremark doctrine (1996), Marchand v. Barnhill (2019), and the Boeing and McDonald's derivative decisions establishing the board's duty to oversee mission-critical risk.
The EU AI Act, the Colorado AI Act, and New York City Local Law 144, among the enacted AI-specific statutes now taking effect across jurisdictions.
SEC, FTC, EEOC, and CFPB guidance and enforcement; Federal Reserve and OCC model-risk expectations; and NAIC, FDA, and FCC sector guidance.
NIST AI RMF, ISO/IEC 42001, ISO/IEC 38507, the OECD AI Principles, COSO, COBIT, and the IIA Global Internal Audit Standards.
The Maturity Model turns the five pillars into something an organization can measure. It expresses the fourteen domains as seventy-six controls and three hundred seven diagnostic statements, and places an organization on a five-level maturity scale across each dimension of the Program.
Each domain resolves into specific controls, and each control into diagnostic statements a board or auditor can test against evidence rather than assertion.
The model places each dimension of the Program on a five-level scale, describing the distance between ad hoc activity and a fully assured, continuously improving oversight program.
The model tailors its diagnostics to whether an organization develops AI systems (Builders) or deploys systems built by others (Buyers). Most organizations are Buyers, and their obligations differ.
When a regulator, a court, or an auditor asks the question, the Program must be able to show the answer. The standard is met when three things are true at once: the Governing Body had what it needed to decide, the Program is real rather than ceremonial, and the chain from policy to practice to assurance is traceable.
A mature Program is not a brake. It is what allows a board to make faster, risk-informed decisions backed by evidence, without sacrificing accountability. That is why oversight governance enables AI adoption rather than constraining it.
The Program overview is open. The full control set, the diagnostic statements, and the Maturity Model are proprietary, and reach practitioners through the Center's co-branded publications and its training for boards, audit executives, and risk leaders.