AI Governance Assessment: Free 28-Question Template (2026)

Most companies find out how their AI governance is holding up at a bad moment, usually when a security review turns up an app nobody approved or a customer asks for evidence that nobody can produce. An AI governance assessment is how you find out first, on your own schedule.

The stakes went up once business teams started building their own tools with vibe coding platforms, since shadow AI spreads faster than any policy document can keep up.

The assessment below is the one we'd hand an IT or security leader: 28 scored questions across seven domains, each mapped to NIST AI RMF and ISO/IEC 42001, plus a scoring guide that tells you what to fix first.

What is an AI governance assessment? The 30-second answer

An AI governance assessment is a structured review of how well your organization controls the AI it builds, buys, and uses, scored against a defined set of questions so you can see where the gaps are and fix the riskiest ones first.

It checks whether your AI governance program works in practice: whether you know what AI is running, who can touch it, what data it sees, how it's tested, and whether you can prove any of that to an auditor.

Three terms get mixed up constantly, so here's how they differ:

What an AI governance assessment should include before deployment

If you're about to put an AI system into production and only have time for the essentials, cover these seven checks. Each one maps to a domain in the full assessment below.

• An inventory entry: the system has an owner, a stated purpose, a list of the data it touches, and a record of where it runs.
• A risk classification: someone has rated its risk level and documented why, including whether it falls into a regulated category.
• Data and privacy review: personal or sensitive data is identified, minimized, and handled under your privacy rules, with a lawful basis where regulations require one.
• Pre-release testing: the system has been tested for accuracy, harmful output, prompt injection, and data leakage, and the results are written down.
• Scoped access: the system and its users run on least-privilege permissions tied to your identity provider and a managed service account.
• Evidence on file: approvals, test results, and the risk rating are stored somewhere an auditor can find them.
• Monitoring and a stop button: logs are on, someone watches them, and there's a named person who can pull the system if it misbehaves.

That list is short on purpose. The gap between "we have a policy" and "we can show you the evidence" is where most programs fall down, and IBM's Cost of a Data Breach 2025 research makes the stakes plain.

Of the 13% of organizations that reported breaches of AI models or applications, 97% said they lacked proper AI access controls. And 63% of breached organizations either had no AI governance policy or were still writing one.

How to run an AI governance assessment in 5 steps

Step 1: Set the scope and the owner

Decide what's in scope: every AI system in the company, one business unit, or one high-risk use case. Name one owner who collects answers and signs off on scores. If you don't have a written AI governance policy yet, note it now, because several questions assume one exists.

Step 2: Gather evidence before you score

Pull the artifacts first: the AI inventory, access-control settings, test reports, data-processing records, and log retention settings. Scoring from memory produces flattering numbers (everyone thinks their team is the careful one).

Step 3: Score each question from 0 to 3

Use the scale in the template and hold yourself to the evidence rule: a 2 or a 3 requires a named artifact you could hand an auditor today. If you can't point to it, score it a 1.

Step 4: Total the domains and read the bands

Add up each domain (maximum 12 points) and the overall score (maximum 84). The interpretation guide below tells you what each band means and which weaknesses to tackle first.

Step 5: Turn the gaps into owned work and re-score

Every question scored 0 or 1 becomes a task with an owner and a due date. Re-run the assessment next quarter, or before any major AI launch, and compare the domain scores.

The AI governance assessment template: 28 questions across 7 domains

Score every question with this scale:

Each question lists the NIST AI Risk Management Framework function it supports (Govern, Map, Measure, or Manage) and the ISO/IEC 42001 clause it relates to.

Domain 1: Inventory and discovery

You can't govern what you can't see, so this domain comes first. Tools for shadow AI discovery help fill it in.

• Q1. Do you keep a single, current inventory of every AI system in use, including vendor tools, internal models, and apps business teams built themselves? (NIST Map · ISO 42001 clause 4)
• Q2. Does every inventory entry name an accountable owner, a business purpose, and the data sources it touches? (NIST Govern · clause 5)
• Q3. Do you have a way to discover AI use that nobody registered, such as network, SSO, expense, or browser-extension reviews? (NIST Map · clause 8)
• Q4. Is the inventory updated automatically whenever a system is deployed, changed, or retired? (NIST Manage · clause 8)

Domain 2: Risk classification

Rate every system so the controls match the stakes. Our guide to AI risk management frameworks covers the methods in depth.

• Q5. Is every AI system assigned a documented risk level using defined criteria (impact on people, data sensitivity, autonomy, and scale)? (NIST Map · clause 6)
• Q6. Do you check each system against regulatory risk categories, such as the EU AI Act's high-risk list, before it launches? (NIST Govern · clause 6)
• Q7. Does a system's risk level decide which approvals, tests, and monitoring it needs? (NIST Manage · clause 6)
• Q8. Do you re-rate a system when its purpose, data, or level of autonomy changes? (NIST Manage · clause 8)

The regulatory question in Q6 has a hard date attached. The EU AI Act sorts AI into four risk levels (unacceptable, high, transparency, and minimal), and the rules for high-risk use cases in sensitive areas like employment and education apply from December 2, 2027.

Domain 3: Data and privacy controls

Most AI risk is data risk wearing a different hat. If you process EU personal data, GDPR-compliant AI practices apply here directly.

• Q9. Do you know which AI systems process personal, confidential, or regulated data, and where that data is stored and processed? (NIST Map · clause 8)
• Q10. Are there rules for what data employees may paste or upload into AI tools, and are those rules technically enforced? (NIST Govern · clause 7)
• Q11. Is training, fine-tuning, and retrieval data reviewed for consent, licensing, quality, and bias before use? (NIST Measure · clause 8)
• Q12. Do vendor contracts state whether your data is used to train their models, where it's hosted, and how long it's kept? (NIST Govern · clause 8)

Domain 4: Model and output testing

Testing is where good intentions meet real prompts.

• Q13. Is every AI system tested for accuracy against its intended use before release, with pass criteria set in advance? (NIST Measure · clause 8)
• Q14. Do you test for harmful or biased output and for prompt injection, jailbreaks, and data leakage? (NIST Measure · clause 8)
• Q15. Is AI-generated code and every AI-built app scanned for security issues (secrets, vulnerable dependencies, insecure data access) before it ships? (NIST Manage · clause 8)
• Q16. Are tests re-run after model updates, prompt changes, or new data sources, and are results kept? (NIST Measure · clause 9)

Domain 5: Access controls

Access is the control the IBM data points to most directly.

• Q17. Do AI systems and AI-built apps authenticate through your identity provider with SSO? (NIST Govern · clause 8)
• Q18. Are permissions role-based and least-privilege, for both the people using AI and the AI systems themselves? (NIST Manage · clause 8)
• Q19. Do AI agents and apps use managed service credentials, with secrets kept out of code and out of personal accounts? (NIST Manage · clause 8)
• Q20. Is access reviewed on a schedule and revoked automatically when someone changes roles or leaves? (NIST Manage · clause 9)

Domain 6: Documentation and evidence

This is the domain auditors care about most, and the one teams most often skip. Our guide to AI governance documentation covers what to keep.

• Q21. Does each AI system have a current record covering purpose, data, risk level, owner, and known limitations (a model card or system card)? (NIST Govern · clause 7)
• Q22. Are approvals, risk ratings, and test results stored in one place tied to the inventory entry? (NIST Govern · clause 7)
• Q23. Can you show who built or changed a system, what changed, and who approved it, for any date in the past year? (NIST Manage · clause 7)
• Q24. Is your AI governance policy published, versioned, and acknowledged by the people it applies to? (NIST Govern · clause 5)

Domain 7: Monitoring and incident response

Governance doesn't end at launch. An AI audit trail is what makes this domain scoreable.

• Q25. Are usage, inputs, outputs, and actions of AI systems logged, with retention that matches your regulatory needs? (NIST Measure · clause 9)
• Q26. Does someone review those logs or alerts on a schedule, with thresholds for drift, errors, and misuse? (NIST Measure · clause 9)
• Q27. Is there an AI incident process that says who can pause or roll back a system, and how fast? (NIST Manage · clause 10)
• Q28. Do incidents and near-misses feed back into risk ratings, tests, and policy updates? (NIST Manage · clause 10)

How to score and read your results

Add each domain (0 to 12) and the total (0 to 84), then find your band:

The total tells you roughly where you stand, and the domain scores tell you what to do next. Two rules keep the results honest:

• Fix the floor first: any domain scoring 4 or below needs attention before you polish a strong one, because one missing domain (usually inventory or access) undoes the rest.
• Inventory and access go first when they're tied: if several domains are weak, start with Domain 1 and Domain 5. Every other control assumes you know what exists and who can reach it.

Our take: most teams land in the Partial band on their first honest pass, and that's useful information. The usual pattern is a written policy (Q24 scores well) with weak discovery (Q3) and no automatic inventory updates (Q4), so the policy governs the systems people remembered to register.

The IBM numbers point the same way. Among organizations with AI governance policies in place, only 34% said they perform regular audits for unsanctioned AI.

AI governance assessment vs. a maturity model

An AI governance assessment is a point-in-time check of specific controls, and a maturity model is a ladder that describes how advanced the whole program is. The assessment tells you that Q18 is a 1, while the maturity model tells you your program sits at level 2 of 5.

They work best together. Run the assessment to find and fix control gaps, then use the domain scores as evidence when you place yourself on the AI governance maturity model. Moving up a maturity level usually means raising two or three domains from Defined (2) to Enforced and evidenced (3).

Both lean on the same frameworks. NIST's AI Risk Management Framework, released January 26, 2023, organizes its core into four functions: Govern, Map, Measure, and Manage.

ISO/IEC 42001, published in December 2023, sets requirements for an AI management system that organizations can certify against.

How Superblocks supports AI governance assessments

Superblocks is the governed enterprise vibe coding platform: business teams build apps with AI, and IT configures the guardrails once. For AI-built internal apps, that turns several assessment domains from manual evidence-gathering into platform defaults.

Inventory and discovery (Domain 1)

The Superblocks MCP server lets IT list applications, manage integrations, deploy apps, and manage access control from AI tools like Claude Code and Cursor. Every app built on the platform is registered as it's created, which answers Q1 and Q4 for that part of your portfolio.

Access controls (Domain 5)

Apps run on SSO and role-based access control that IT sets up once, and imported prototypes from Claude, Lovable, and Replit move off hardcoded credentials onto platform-managed authentication and managed integration credentials.

Documentation, evidence, and monitoring (Domains 6 and 7)

Audit logs give admins a searchable record of activity across the organization, filterable by actor, event type, resource, and severity. That's the evidence Q23 and Q25 ask for.

Testing and standards (Domain 4)

The Security Agent reviews each app before publish for issues like hardcoded secrets, sensitive-data leaks, and endpoints that expose too much, and organization knowledge applies your design and coding standards to every build.

Data residency (Domain 3)

Superblocks offers Cloud, Hybrid, and Cloud-Prem deployment. With Cloud-Prem, the full platform, including AI inference, runs inside your own AWS environment.

Audit logs, SSO, VPC deployment, the Admin MCP, and security agents are part of the Enterprise plan. RBAC is available on Teams and up.

Score it now, then score it again next quarter

The first AI governance assessment is mostly a discovery exercise. You'll find systems nobody registered and controls that exist on paper only, and the useful move is to write those down without flinching and assign every 0 or 1 an owner.

The second pass is where it gets valuable, because the domain scores start to show whether the work is sticking. The domains that stay stubbornly low are usually the ones that depend on people remembering to do something, and those are the ones worth moving onto a platform that does it by default.

With Superblocks, teams can:

• Register every AI-built internal app in one governed place as it's created
• List and administer apps, integrations, and access through the Superblocks MCP
• Run every app on SSO and role-based access controls IT configures once
• Search audit logs by actor, event, resource, and severity when evidence is due
• Scan apps for secrets and data-exposure issues with the Security Agent before publish
• Apply design and coding standards automatically through organization knowledge
• Keep production data and AI inference inside your own cloud with Cloud-Prem

When the controls live in the platform, the next assessment is mostly an afternoon of exporting evidence.

Book a demo to see how Superblocks handles the inventory, access, and evidence domains for AI-built apps.

Frequently asked questions

How often should you run an AI governance assessment?

Run an AI governance assessment at least quarterly and before any major AI launch or regulatory deadline. Quarterly re-scoring shows whether fixes are sticking, and pre-launch checks catch new systems before they reach production.

What is the difference between an AI governance assessment and an AI audit?

An AI governance assessment is an internal review that finds control gaps, while an AI audit independently tests whether controls work and are backed by evidence. Most teams run assessments first so the audit has fewer surprises.

What are the six pillars of AI governance?

The six pillars of AI governance usually cited are accountability, transparency, fairness, privacy and security, safety and reliability, and human oversight. Frameworks group them differently, so map your assessment questions to NIST AI RMF or ISO/IEC 42001 for a defensible structure.

Who should run an AI governance assessment?

An AI governance assessment is usually owned by IT, security, or a risk and compliance lead, with input from legal, data, and the business owners of each AI system. One named owner should collect evidence and sign off on scores.

Is there a free AI governance assessment template?

Yes, the 28-question template in this guide is free to use. It covers seven domains, scores each question from 0 to 3, and maps every question to NIST AI RMF and ISO/IEC 42001.