- CIO
- CISO
- CEO
How to Build an AI Governance Framework Your Team Can Actually Run
THE SHORT ANSWER
An AI governance framework your team can run comes down to four things: a named owner for every AI use, approval rules that match the risk, one inventory of what is approved and what it can reach, and a rule to review again when anything changes. Use NIST's AI Risk Management Framework for structure and keep the rules short.
On this page
A vendor adds AI to an application you already use. The new feature can read customer records. Who reviews that change, who can approve it, and who can turn it off while the review is underway?
An AI governance framework gives your team a consistent way to answer those questions, for every AI tool, feature and agent, without starting from scratch each time.
What should an AI governance framework cover?
It applies to three kinds of AI:
- standalone AI tools,
- AI features built into software you already use,
- agents that connect to company systems.
Use it when you buy or build something new, and when a vendor changes a product you already approved.
For every use case, record a business owner, what the AI may access and do, who can approve or suspend it, and when the decision must be reviewed.
Who owns the decision?
- The business owner is accountable for the outcome and defines how success will be measured.
- The CIO or IT lead confirms what the AI can access and whether the proposed limits can be enforced.
- The CISO evaluates the security exposure.
- Privacy or Legal joins when the data or the use warrants it.
Name the approver for each risk level, and the person authorized to accept exceptions.
How should approval rules change with risk?
Start with a short intake form that captures the purpose, users, data, provider, connected systems and the actions the AI can take. Then route by risk:
| Risk tier | Typical examples | Who approves | Evidence before approval | When to review again |
|---|---|---|---|---|
| Low | An approved tool used within policy, with no sensitive data | Business owner, through standard approval | Completed intake form | Yearly, or when it changes |
| Medium | Use of internal confidential data, or a connection to company systems | Business owner and IT | Access limits tested, data rules agreed | Every six months, or when it changes |
| High | Customer or other sensitive data, customer-facing output, or automated actions | Business owner, IT and the CISO, with Privacy or Legal where needed | Tested controls, human approval for high-impact actions, a working way to stop it | Before any expansion, and whenever it changes |
Give every pilot a defined scope and an end date. When a request falls outside policy, send it to the designated risk approver, and record the reason, safeguards, scope, expiry and who can revoke access.
How do you keep track after approval?
Keep one inventory with each use case's owner, approved scope, permissions, sign-offs and review dates. Reopen approval when the AI gains new data, permissions or actions, or when its provider or data handling changes. Assign someone who can pause the feature or revoke access while it is reviewed.
What does this look like for a real request?
The support team wants to use AI to draft replies from customer records. Approve a limited pilot if IT can restrict access to the records the team needs and staff review each reply before it is sent. Access to the full customer database goes through the exception process. Letting the AI send replies on its own triggers another review.
Which standards should you use?
| Standard | What it is | Certifiable? |
|---|---|---|
| NIST AI Risk Management Framework 1.0 | Released January 26, 2023 and "intended for voluntary use." It organizes the work into four functions: Govern, Map, Measure and Manage. NIST says it "is being revised as part of the White House AI Action Plan" | No |
| NIST AI 600-1, Generative AI Profile | Released July 26, 2024, with suggested actions for generative AI, including third-party AI | No |
| ISO/IEC 42001:2023 | An international standard for an AI management system, published December 18, 2023 | Yes. Organizations can be certified against its requirements |
Most companies can start with NIST's structure and short internal rules. Consider ISO/IEC 42001 certification when customers, partners or regulators ask for formal proof.
Three checks you can run
| Check | What a pass looks like | What a fail looks like |
|---|---|---|
| Does every AI tool, feature and agent in use have a named business owner? | One inventory with an owner for each entry | AI tools found only when someone asks |
| Can you say who approves a high-risk AI use, and what evidence they need first? | Written approval rules by risk tier | Approval depends on who you ask |
| Is there a rule that reopens approval when an AI tool gains new data, permissions or actions? | A review trigger that application owners know about | Approval is a one-time event |
Questions we hear
Do we need ISO/IEC 42001 certification?
Not unless customers, partners or regulators ask for it. ISO/IEC 42001 is a certifiable management system standard, while NIST's AI Risk Management Framework is voluntary and can't be certified against. Many companies start with NIST-based internal rules and consider certification later, when there is a clear business reason.
Who should own AI governance?
Ownership works best when it is split by role. The business owner is accountable for each use case and its results. IT confirms what the AI can access and whether limits can be enforced. Security reviews the exposure. Privacy or Legal joins when sensitive data is involved. One person should coordinate the inventory.
What counts as an AI use case that needs approval?
Standalone AI tools, AI features added to software you already use, and agents that connect to company systems. The last two are the ones most often missed, because they arrive through software updates or are built by teams without going through procurement.
What should happen when a vendor switches on a new AI feature?
Treat it as a change to an approved use case. Reopen approval if the feature gains access to new data, can take new actions, or sends data to a new provider. Name someone who can pause the feature while it is reviewed.
Sources
- NIST, AI Risk Management Framework (opens in a new tab)
- NIST AI 100-1, AI Risk Management Framework 1.0 (January 2023) (opens in a new tab)
- NIST AI 600-1, Generative Artificial Intelligence Profile (July 2024) (opens in a new tab)
- IEC webstore, ISO/IEC 42001:2023 (opens in a new tab)
- Schellman, ISO/IEC 42001 requirements (opens in a new tab)
