← Back to the Program

I Agile Governance The Constitution

The mandate, the structure, and the authority that give the program its standing. Without this pillar there is no program to review.

01. AI Governance Program and Policy Framework

The mandate for the program itself. This domain governs whether AI oversight exists as a matter of record: a charter, a resolution, or an amendment to an existing charter that names AI as something the board oversees, and the policy set that sits beneath it. The mandate also fixes the program's scope and grain: which parts of the organization and which kinds of AI use it reaches, and at what level of detail, sized to this organization rather than copied from a template. That scope is defined, documented, and visible to the board. The Governing Body owns the mandate. One executive or one constituted body is named as accountable for the program as a whole. The board question: can we produce the document that says we govern AI, what it covers and at what grain, and the date we last reviewed it? Domain 01 establishes that the program exists, what it reaches, and who owns it. Domain 02 establishes what that owner is allowed to decide and what they have to decide with.

02. AI Governance Structure, Oversight, and Resources

The authority and the resources to act. This domain governs the decision rights of the body that oversees AI: what it approves, what it can stop, and what it must escalate. It also governs whether that body is funded and staffed to the size of its mandate, and whether anyone has assessed that fit rather than assumed it. The accountable function owns the structure and the reporting line. The Governing Body confirms that the budget and staff match the mandate, because an oversight body that can only receive updates is not an oversight body. The board question: how does management assess whether the oversight function has the resources to develop and adopt what the mandate requires, and has the board seen that assessment? Domain 02 is about the standing structure. Domain 13 is about what happens inside that structure when a specific risk crosses a threshold.

03. Governance Program Assurance and Continuous Learning

The program reviews itself, as designed. This domain governs whether the oversight program is a subject of assurance in its own right: reviewed on a defined cycle by internal audit or an independent party, and changed as a result of what the review finds, what the organization learns from incidents, and what regulators publish. A named function owns the review cycle and the follow-up. The board question: has the AI oversight program itself been audited, or is it on the audit plan, and is the cadence right for how fast our use of AI is changing? Domain 03 tests the program as it was designed. Domain 14 tests whether the escalation path works in use. One asks whether the structure is sound; the other asks whether the brakes engage.

II Risk-Informed System The Guardrails

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 pillar turns on whether a defined methodology exists for setting tolerance and identifying risk. The Governing Body requires that methodology and approves the tolerance. The decisions made inside it belong to management.

04. AI Risk Methodology, Scope, and Tolerance

What is in scope, and how much risk is acceptable. This domain governs the methodology every other domain depends on: how AI risk is identified, how tolerance is set, which AI activities the program reaches, and whether that methodology has been approved as appropriate for this organization. Scope has to name the hard cases: third-party AI, AI embedded in enterprise products, employee use of generative AI, and systems that act on their own. Tolerance has to be specific enough that management can tell when an activity exceeds it. A general commitment to responsible AI does not meet that test. The Governing Body approves the appetite and confirms the methodology fits the organization. A named function owns the methodology, the scope decision, and the running of use cases through it. The board question: are the AI use cases that carry risk being worked through the methodology, and which ones are not? Domain 04 sets the boundary. Domain 06 holds individual systems to the standard inside it.

05. Risk Intelligence and Threat Landscape

Knowing what is changing outside the organization. This domain governs whether a standing capability exists to learn about emerging AI risks, regulatory developments, and incidents elsewhere, and to decide which of them matter here. The point is not a news feed. The point is a defined intelligence function with an owner, a method for deciding relevance, and a cadence for reaching the oversight body. A named function is accountable for it. The board question: is there a standing AI threat and regulatory intelligence capability, with an owner and a cadence to the oversight body, or do we learn about outside events by accident? Domain 05 looks outward at what is changing. Domain 09 looks inward at whether our own systems are protected against it.

III AI Trust and Assurance The Evidence

The standards AI systems must demonstrably meet, and the assurance that reaches the oversight body to prove they meet them.

06. AI Model Risk and Agentic Lifecycle Oversight

Standards for AI systems across their life, including systems that act. This domain governs what every AI system in scope must demonstrably meet before it is deployed and while it is in use, and it requires a current inventory of those systems so that the standard has something to apply to. Agentic systems, which take actions rather than produce outputs, are inside this domain, not a special case outside it. The Governing Body requires that the standards exist and that they are approved. Management sets them and holds systems to them. Ownership is assigned for every system in the inventory, and a function is named to confirm the inventory is complete. The oversight body receives attested reporting on whether systems meet the standard, including degradation and incidents. The board question: how many AI systems do we run, who owns each one, and which ones are not meeting the standard today? Domain 06 governs every system in scope across its life. Domain 12 governs the terms on which outside systems enter that scope.

Buyers: the standard applies to the systems you deploy, not only the ones you build, and the inventory of them is yours to keep.

07. AI Data Governance Oversight

What the organization requires of the data AI learns from and acts on. This domain governs the requirements placed on the data behind AI systems: where it came from, whether it is fit for the purpose, whether its use is lawful, and how it is protected. It does not duplicate the organization's existing data governance. It states what that governance has to cover once AI is learning from the data and acting on it. A named function is accountable for data governance as it applies to AI, separate from the teams that build the systems. The board question: do we know what data our AI systems were built on, and whether we had the right to use it? Domain 07 governs the data. Domain 06 governs the system that uses it.

Buyers: the requirement reaches the data you feed a vendor system and the data its outputs act on.

08. AI Transparency, Explainability, and Human Oversight

When a decision must be explained, and when a person must be able to intervene. This domain governs the organization's own rules for AI-influenced decisions: which decisions must be explainable, to whom, and when a person must be able to review, intervene, or reverse the outcome. The rule is proportionate to the risk; not every decision needs a human in the loop, and the program says which do. Where it does, the human role is a documented procedure, not a habit. Explainable means the decision process and the information behind it can be set out; it does not require that the model itself be interpretable. For each higher-risk use, a named role is accountable for human oversight of the system's outputs and actions. The board question: are our high-risk AI decision processes documented, with the human role defined as a procedure, and can each decision be explained even where the model cannot be interpreted? Domain 08 governs the human role in decisions the system makes. Domain 13 governs what happens when the system, or the human, finds a risk that has to go up.

09. AI Security and Resilience Assurance

Assurance that AI systems are secured and that the organization can carry on if one fails. This domain governs two requirements: that AI systems are protected against misuse, manipulation, and failure, and that the organization can keep operating if a material AI system fails or must be withdrawn. The oversight concern is the seam. AI security and AI resilience sit between the existing security function and the existing business continuity function, and this domain requires that accountability be assigned and reconciled so that nothing falls between them. Resilience is measured, not assumed: indicators exist for the material systems, with thresholds, and they move as the AI environment moves. Independent assurance results reach the oversight body with the material findings and their disposition. The board question: are resilience indicators for our material AI systems established, with thresholds, and adjusted at the pace the AI environment changes? Domain 09 is about whether the system survives attack or failure. Domain 11 is about whether the business gets what it was promised from it.

IV Risk-Based Strategy and Execution The Strategic Alignment

The link between oversight and the decisions that commit capital, accept risk, and bring AI in from outside. This is where the balance of risk and reward is struck, and where it is revisited at the pace this environment moves. The board sees the balance. Management strikes it.

10. Risk-Informed Strategy, Resources, and Organizational Readiness

Oversight before the commitment, not after it. This domain governs whether AI oversight has a seat at the point where strategy and investment decisions are made, rather than arriving at deployment when the money is spent and the direction is set. It also governs readiness: whether the organization has the capacity, skills, and resources to govern the AI it plans to adopt, before it adopts it. A named executive is accountable for confirming that readiness. Records of major AI investment decisions show that oversight considerations were presented and weighed. The board question: taken together, are our AI investments inside the appetite and tolerance we set for risk and reward, and can the board see that fit without stepping into the investment decisions themselves? Domain 10 governs the decision to commit. Domain 11 governs what happened after the commitment was made.

11. AI Value Realization and Operational Resilience Oversight

Whether AI is delivering what was approved, within the risk that was accepted. This domain governs the closing of the loop. At approval, the expected value and the accepted risk of a material AI initiative are stated. Afterward, the program requires reporting against both, so the oversight body can see whether the initiative delivered and whether it stayed inside the risk that was agreed. A named function owns that reporting. The oversight body has received at least one such report and can show what it decided as a result. The board question: of the AI initiatives we approved, which delivered, which did not, and what did we do about the ones that did not? Domain 11 governs performance against the approval. Domain 09 governs whether the system survives attack or failure.

12. Third-Party AI and Supply Chain Governance

AI the organization consumes rather than builds. Most organizations do not build the AI they rely on. They license it, inherit it inside vendor platforms, or consume it from foundation model providers. This domain governs all of it as in scope, with requirements set for each kind, and it requires that the organization understand what a supplier failure or withdrawal would mean for its own operations and whether that exposure fits the risk and reward it accepted. Accountability for AI risk that originates with a supplier is assigned inside the organization. A vendor can be the source of the risk. It cannot be the owner of it. The inventory identifies third-party AI dependencies, and the oversight body has visibility into the material ones. The board question: do we understand the resilience implications of the AI we depend on from outside, does that exposure fit the risk and reward profile we accepted, and who inside is accountable for knowing? Domain 12 governs the terms on which outside AI enters scope. Domain 06 governs every system, outside or inside, once it is in.

Buyers: this is where most of your AI risk originates. Accountability for it stays inside the organization, not with the supplier.

V Escalation and Disclosure The Voice

How a critical risk reaches the people who can decide, how the organization discloses it, and how it knows the path works. A tolerance is not a no. It is a decision point. When a risk falls outside tolerance, the question is whether the decision reaches the people with the authority to make it, in time, and whether the organization meets its disclosure responsibilities once it does.

13. AI Risk Escalation and Disclosure Protocols

Critical risks bypass silos and reach decision makers. This domain governs the path a serious AI risk or incident takes upward: the thresholds that trigger escalation from the first line to the second and from management to the Governing Body, the timing attached to each, and the speed at which a material risk is recognized as material in the first place. It also governs disclosure, which arises for three reasons: because a risk is material and disclosure is an obligation, because it meets a threshold a regulator demands, or because the public's trust depends on hearing it from the organization first. A named role decides, in a specific event, whether a threshold has been met and makes the escalation. A named role owns each disclosure obligation. The board question: when a material AI risk emerges, how fast do we know, and do we know which of our disclosure obligations it triggers? Domain 13 governs the path up. Domain 14 governs whether the path works.

14. Validation of Escalation and Governance Effectiveness

Testing the emergency brakes, as used. This domain governs whether escalation and oversight actually function: whether risk information reaches decision makers intact, in time, and in a form that allows a decision. The test is independent of the first line, and its results go to the Governing Body as a report on effectiveness, not a summary of incidents. A function independent of the first line is accountable for the testing and the reporting. The board question: have the AI risk escalation protocols been tested, and did they work? Domain 14 tests the escalation path in use. Domain 03 reviews the program as designed. Both are assurance; they look at different things.

← Back to the Program