AI governance framework: what it is and how to build one

The four things every framework has to answer

Frameworks differ in vocabulary but converge on the same questions. If a framework cannot answer these, it is a statement of values rather than a governance structure.

• Accountability — which named person owns this system, and who approves changes to its authority.
• Authority — what the system is permitted to access and do, and under which conditions.
• Enforcement — what actually stops a prohibited action, as distinct from what discourages it.
• Evidence — what record proves, after the fact, that the control operated.

NIST AI RMF, ISO/IEC 42001 and the EU AI Act compared

Three published references cover most enterprise requirements, and they compose rather than compete. They are different kinds of document: one is voluntary guidance, one is a standard you can be certified against, and one is law. The primary source for each is linked in the last row.

How the three fit together

Most organizations use NIST AI RMF to structure the work, ISO/IEC 42001 to make it certifiable, and the EU AI Act to determine which obligations are mandatory rather than advisable. Using one does not discharge the others: certifying to ISO/IEC 42001 does not by itself satisfy the AI Act.

For the AI Act dates, see the EU AI Act timeline.

Where frameworks usually fail

The common failure is a framework that terminates in documentation. Controls are described, owners are listed in a spreadsheet, and nothing in the running system enforces any of it. The gap becomes visible the first time an auditor asks not whether a policy exists, but which specific action it prevented, and when.

The second failure is scope. A framework written for models governs training data, evaluation and bias. It says little about an autonomous agent that holds credentials, calls tools, and takes actions in production systems — where the risk is not what the model says but what the software does.

From framework to enforced control

Turning a framework into something operational means binding each element to a mechanism: accountability to an owner recorded against a registered agent; authority to a policy evaluated at the moment of action; enforcement to a decision point that can actually deny; evidence to a record generated by that decision rather than reconstructed later.

That is the difference between a governance programme that can describe its controls and one that can demonstrate them.

A 7-step rollout

Start with the systems, not the standard. This is our suggested order, not a sequence any of the three frameworks requires. Each step produces an artifact the next one consumes, which prevents a familiar failure: mapping controls to NIST, ISO or the EU AI Act before you know what is in scope.

1. Inventory every AI system and agent. List what is in use, including tools bought by individual teams and agents that hold credentials. Nothing later can be scoped against systems nobody has listed.
2. Name an accountable owner. One named person per system, and a named body that approves changes to its authority.
3. Classify use, data and affected people. Record what the system is for, which data it touches and who its decisions affect. This is the input to any risk tier.
4. Record authority and dependencies. Write down which tools, data and actions the system may use, and which models, vendors and other agents it depends on.
5. Choose controls for the actual risk. Pick controls to match the tier and the authority recorded above, then map them to NIST AI RMF, ISO/IEC 42001 or the EU AI Act as needed.
6. Require an approval decision with evidence. A named approver signs off before go-live, and the decision is stored as a record.
7. Monitor, and reassess on a date. Review when a system gains a new tool, data class, model, vendor or user group, and on a fixed calendar date regardless.

The minimum operating artifacts

A small organization does not need a large governance office, but it does need a policy, an AI system inventory, a risk register, one assessment per material system, approval records, incident and change logs, and a review calendar. The documents matter because they assign responsibility and preserve decisions; runtime evidence matters because it shows whether those decisions operated.

Common questions

What are the main AI governance frameworks?

The three most referenced are the NIST AI Risk Management Framework (a voluntary framework), ISO/IEC 42001 (a certifiable AI management-system standard) and the EU AI Act (binding regulation for AI placed on or used in the EU). They compose rather than compete: NIST AI RMF structures the risk work, ISO/IEC 42001 makes the management system certifiable, and the EU AI Act decides which obligations are mandatory.

Is there a free AI governance framework template?

Yes. Agent Trust Cloud offers a free, editable AI governance framework template in Word format, mapped to NIST AI RMF and ISO/IEC 42001. An email address unlocks the download; there is no payment. Free AI governance framework template

Sources

• NIST AI Risk Management Framework (NIST AI 100-1) (primary source)
• ISO/IEC 42001 (iso.org) (primary source)
• Regulation (EU) 2024/1689, the AI Act (EUR-Lex) (primary source)