Governance first: audits, frameworks and vendor disclosures
How I’d set up AI governance at a company, audit it against NIST and ISO, and hold vendors to written AI and data disclosures.
Governance comes first
Everyone wants to talk about what AI can do. I care more about who owns it and who approved it. And when something goes wrong, who gets the call? A company with clear answers can move faster, because nobody has to stop and ask.
I studied this in Data and Technology Governance at ASU, and I see it every day running the AI portfolio at work. Every project already has a portfolio owner and an executive stakeholder. I’d apply the same idea to AI across the whole company.
Establish it before you scale
- Own
Name an executive sponsor and a cross-functional council: legal, compliance, risk, IT and security, data and the business.
- Inventory
One list of every AI system, whether it was built, bought or is embedded in a vendor’s product.
- Tier
Classify each use by risk: what it can see, what it can do, and what a wrong answer would cost.
- Audit
Check regularly that what’s running still meets the frameworks the company has committed to.
- Manage vendors
Hold every third party to written AI and data disclosures, and keep checking them.
The order matters. You can’t audit a system that isn’t on the list, and you can’t hold a vendor to rules you never wrote down.
Audits against real frameworks
Approving a policy once and filing it doesn’t protect anyone. Audits tell you whether the system running today still matches the one that was approved. I’d anchor them to recognized frameworks, so the questions aren’t invented each time:
| Framework | What the audit checks |
|---|---|
| NIST AI Risk Management Framework | Its four functions: Govern (owners, policy, accountability), Map (context and impact), Measure (testing and monitoring) and Manage (response and treatment of risk). |
| NIST Generative AI Profile (AI 600-1) | Risks specific to generative AI, such as made-up answers, data leakage and harmful output, and the controls for each. |
| ISO/IEC 42001 | Whether the company runs a documented AI management system: roles, objectives, controls and continual improvement. |
| Confidentiality, integrity and availability | The data side: who can see it, whether it can be trusted, and whether it’s there when it’s needed. |
A good audit tests whether the controls actually work. That includes human review. If someone clicks “approve” on output they didn’t have time to read, the review step only exists on paper.
Vendor management runs through governance
Most of the AI a company uses, it doesn’t build. It arrives inside software it already pays for, and vendors add AI features all the time, sometimes without much notice. So I’d put vendor management inside the governance program.
Every contract with a third party should spell out its AI and data disclosures, and those disclosures have to reach past the vendor itself to the companies it relies on:
| Disclosure | What the contract should require |
|---|---|
| Where AI is used | Which features use AI, which models power them, and notice before new AI features are switched on. |
| Training on your data | Whether your data, prompts or outputs are ever used to train or improve their models, with a default of no. |
| Retention and deletion | How long prompts, outputs and logs are kept, and proof of deletion at the end of the contract. |
| Subprocessors | Every fourth party that touches your data, including the AI model providers behind the vendor, with notice of changes. |
| Model changes | Notice when the underlying model or provider changes, since a new model can behave differently. |
| Audit and exit rights | The right to review their controls, and a clean way out that returns your data. |
Vendors get reassessed after signing too, whenever they ship a major release, change models or add subprocessors. When I run a structured pilot of a new AI platform, these are the questions that come before any feature demo.
Where to start
- Name the owners. Every AI system in use gets a business owner and a risk tier this quarter.
- Build the inventory. Include AI built into vendor products and the tools employees started using on their own.
- Pick the framework. Use the NIST AI Risk Management Framework as the backbone, so audits ask the same questions every time.
- Review the top vendors. Start with the vendors that touch the most sensitive data, and check their contracts against the disclosures above.
- Schedule the audits. Put audits on the calendar, plus triggers for a re-review: a new model, new data or a new kind of decision.