TL;DR
An AI governance assessment checks whether your AI controls actually work, using evidence instead of policy documents. You run it in seven steps. First set the scope and owners, then build an inventory of every AI system. Next, collect evidence and answer a scored worksheet across six governance domains. Finally, write findings and turn each gap into a backlog item with an owner and a deadline. Repeat it on a risk-based schedule and whenever a system, vendor or regulation changes.
Key Takeaways An AI governance assessment tests whether controls operate in practice, so it needs records, logs and samples rather than written policy alone. The inventory comes first, because you cannot assess AI systems that nobody has registered, including vendor features and shadow AI. A 24-question worksheet across six domains covers accountability, inventory, lifecycle, monitoring, third-party AI and agent oversight. Missing evidence scores zero, and one failed critical control should block a domain from passing however high its average looks. The output that matters is a remediation backlog with owners, deadlines and evidence-based closure, not the score itself. Self-assessments suit routine checks, while independent reviewers add weight before regulators, boards or certification audits. Watch on YouTube
AI Audit Framework: EU AI Act, NIST, ISO 42001 & the 6-Step Blueprint
A walkthrough of how the EU AI Act, the NIST framework and the ISO management-system standard fit together, plus a six-step blueprint for an audit you can pass.
Could You Prove Your AI Controls Work by Friday?
Suppose your audit committee asks on Monday for three things. They want a list of every AI system the company runs, the person who owns each one, and proof that its controls worked last quarter. How long would it take your team to answer, and how much of the answer would be a guess?
For many companies, the honest answer is a guess. IBM’s 2025 Cost of a Data Breach findings show that 63% of breached organizations had no AI governance policy or were still writing one. Even among those with a policy, only 34% regularly audited for unsanctioned AI.
Yet a policy on a shared drive does not show that anyone follows it. That gap, between what is written and what actually happens, is exactly what the assessment measures.
What Is an AI Governance Assessment?
An AI governance assessment is a structured review of whether an organization’s AI controls exist, operate and produce evidence. It also covers people, process and records across every AI system in scope. After that, each control gets a score, each gap gets a written finding, and the findings become a funded plan.
It sits inside a wider AI governance program and borrows its control list from whichever AI governance framework you follow. Several related reviews often get confused with it, so the table below separates them by the question each one answers.
Review Question it answers Typical output AI governance assessment Do our governance controls operate, and where are the gaps? Scored worksheet, findings, remediation backlog AI governance maturity model How developed is our governance program overall? A maturity level per domain AI readiness assessment Are we ready to adopt AI at scale? Readiness score across data, skills and platforms AI audit Does a system or program meet a defined standard? Formal audit opinion, often independent AI security assessment Can attackers misuse or break our AI systems? Threat and vulnerability findings
The maturity model and the assessment work as a pair. The assessment gathers the evidence and scores the controls, then the maturity model turns those results into a level. So keep the two separate, and run the assessment first.
How to Run an AI Governance Assessment in Seven Steps
The procedure follows the same arc whether you assess one business unit or the whole enterprise. You scope it, inventory the AI estate and gather evidence, then score the worksheet, write findings and build the backlog. The first three steps decide whether the scores mean anything, so they get the most attention here.
Step 1: Set the Scope, Sponsor and Reference Standards
Start with a one-page charter. It names the business units, AI systems and environments in scope, the executive sponsor, the assessment lead and the control owners who must supply evidence. It also states what the assessment will not do, such as technical penetration testing, which belongs to AI red teaming .
For reference standards, most teams map their controls to the NIST AI Risk Management Framework , which NIST released on January 26, 2023 for voluntary use. For GenAI systems, they also add its Generative AI Profile (NIST-AI-600-1). Teams heading toward certification also map to ISO/IEC 42001:2023 , the standard for an AI management system.
Our NIST AI RMF Playbook guide turns that framework into concrete actions. An internal assessment is not a certificate, however, so state that limit in the charter too.
Step 2: Build the AI Inventory Before You Score Anything
NIST’s own core text puts the inventory early. Its GOVERN 1.6 subcategory says mechanisms should be in place to inventory AI systems, resourced according to risk priorities. Without that list, every later score covers only the systems somebody remembered.
Next, search beyond the projects IT already knows about. That means vendor features switched on inside SaaS tools, internal copilots, generative AI used through browser accounts and AI agents that can call tools. Expense reports, SSO logs and network traffic usually surface shadow AI that never went through intake.
After that, record the same minimum fields for every system. Use name, purpose, business owner, technical owner, vendor or model, data used, users affected, deployment status and risk tier.
Then classify each system by impact, because a tool that scores loan applicants needs deeper review than one that drafts meeting notes.
Step 3: Collect Evidence That Controls Actually Operate
Send the evidence request list before the first interview, because control owners need time to pull records. Ask for two kinds of proof, because each answers a different question. Policy evidence shows a control is designed, while operational evidence shows it ran, such as approval tickets, review minutes, monitoring alerts and access logs.
Regulation also makes operational records more than a nice-to-have. Under Article 26 of the EU AI Act , deployers of high-risk systems must monitor their operation. They must also keep the logs those systems generate for at least six months, unless other law says otherwise.
Then test a sample instead of trusting a summary. Pick three systems and trace each one from approval to its latest monitoring report, noting how recent every record is. If a control owner cannot produce a record, log it as missing rather than taking a verbal assurance.
Step 4: The 24-Question AI Governance Assessment Worksheet
The worksheet uses the same six domains as our maturity model, with four questions each. Every question asks whether a control is demonstrated, not whether it is planned. Adapt the wording to your own policy names, but keep the evidence column, because that column turns opinion into a score.
Domains 1 to 3: Policy, Inventory and Lifecycle
Domain Assessment questions Evidence to request 1. Accountability and policy Q1 Is there an approved AI policy with a named owner and review date? Q2 Are business, technical, risk and compliance roles assigned per system? Q3 Is there a process to approve AI use, record exceptions and escalate? Q4 Do staff who build or use AI get role-based AI literacy training? Signed policy, RACI chart, exception log, training records 2. AI inventory and risk tiering Q5 Is there one reconciled inventory of AI systems in scope? Q6 Does each entry list owners, purpose and status? Q7 Is an approved risk-tiering method applied and recorded? Q8 Are unregistered systems found and resolved through a defined process? Inventory export, tiering rubric, discovery reports 3. Lifecycle controls Q9 Are risk and impact reviews completed before production? Q10 Do changes to models, prompts, tools or data trigger re-approval? Q11 Are testing and deployment records kept? Q12 Is there a process to retire systems and close their access? Approval tickets, change records, test reports, retirement checklist
These first three domains test whether governance exists on paper and reaches every system before launch.
Checklist
AI Governance Checklist
A practical checklist of the policies, roles, inventory and monitoring controls an AI governance program needs, ready to use beside your assessment worksheet.
Get the Checklist → Domains 4 to 6: Monitoring, Vendors and Agents
Domain Assessment questions Evidence to request 4. Monitoring and evidence Q13 Are monitoring rules set by risk tier? Q14 Are incidents and control failures logged and tracked? Q15 Do owners review monitoring results on a schedule? Q16 Can you retrieve the full record for any sampled system within a day? Dashboards, incident register, review minutes, log samples 5. Third-party AI Q17 Are vendor AI systems in the inventory? Q18 Are vendor risks, data-use terms and contracts reviewed before approval? Q19 Do new vendor AI features trigger reassessment? Q20 Are exit and escalation duties documented for critical AI vendors? Vendor questionnaires, contract clauses, reassessment log 6. Agent oversight Q21 Is every AI agent registered with an owner and its permitted actions? Q22 Are autonomy limits and human approval points documented and enforced? Q23 Are agent actions and tool calls kept in an audit trail? Q24 Are agent permissions and the emergency stop tested on a schedule? Agent registry, permission matrix, action logs, kill-switch test record
How the Domains Map to NIST and the EU AI Act
Each domain traces back to published guidance. For instance, accountability follows GOVERN 2.1 on documented roles, and Q4 reflects the AI literacy duty in Article 4 of the EU AI Act. Third-party questions follow GOVERN 6.1 on supply chain risk, while the monitoring domain follows MANAGE 4.1 on post-deployment monitoring plans.
Agent oversight earns its own domain because agents take actions on real systems. IBM found that 97% of organizations reporting an AI model or application breach lacked proper AI access controls. That is why Q21 to Q24 lean on permissions and logs, which our guides to agentic AI governance and human-in-the-loop controls for agents also cover.
Each worksheet row should capture the question ID, applicability, evidence link, score and evidence confidence. It should also record the gap, its risk priority, the owner and the recommended action. Keep those fields in one spreadsheet or GRC tool, so a reviewer can follow any score back to its record.
Step 5: Score the Results Without Hiding Critical Gaps
Use a simple three-point scale. A score of 0 means not demonstrated, 1 means partly demonstrated, and 2 means demonstrated with sufficient evidence. A wider scale invites debate about the difference between a 3 and a 4, while a narrow one keeps attention on the evidence.
Next, calculate each domain as points achieved divided by points available. Mark a question N/A only with a written reason, then drop it from the denominator. Missing evidence is never N/A, though, and it scores 0 because an assessor cannot credit a control nobody can show.
Averages can hide a serious hole, so flag a few critical controls, such as the inventory, pre-production review and agent kill switch. If any of them scores 0, the domain fails, whatever its percentage says. That way, a 90% domain score can never mask an agent nobody can switch off.
Also record evidence confidence beside each score, as high, medium or low. Then hold a factual review where control owners can correct errors and add records, but cannot overrule a finding without evidence. The scores feed the maturity levels later, so keep the raw worksheet unchanged.
Steps 6 and 7: Write Findings and Build the Remediation Backlog
A finding needs five parts, namely the condition, the evidence, the business or regulatory impact, the accountable owner and the recommended action. For instance, a finding might read “Vendor AI features are not in the inventory, so 14 SaaS tools process customer data without a risk review.” “Improve vendor oversight” is not a finding.
Then group findings by domain and by risk priority, and separate enterprise-wide gaps from single-system failures. A missing policy affects every system at once, while one unmonitored chatbot is a local fix. Leaders read the first group, while engineering teams work from the second.
Next, turn each finding into a backlog item with an owner, a due date and a closure test. Split the work into three horizons.
In the first 30 days , contain the critical gaps, such as pausing an unregistered agent or restricting its permissions.By day 60 , fix the process gaps, such as adding vendor AI to procurement intake.By day 90 , finish the structural work, such as tooling for monitoring or a central registry.Finally, close an item only when new evidence proves the control now runs. For example, a rewritten policy closes nothing until a sample shows people following it. Funding and operating-model choices come next, and our guide to AI governance strategy covers them, along with the playbook on AI governance best practices .
Self-Assessment or Third-Party AI Governance Assessment?
Both approaches use the same worksheet, so the choice comes down to who needs to trust the result. Internal teams know the systems, so they can run checks often. Independent reviewers, on the other hand, bring distance, which matters when the audience is a regulator, a board or a certification body.
Factor Self-assessment Third-party assessment Best for Routine checks, a first baseline, single business units Board reporting, regulated systems, pre-certification readiness Independence Low, since owners may assess their own controls High, with findings that carry weight with auditors Speed and cost Faster to start, mostly internal time Needs procurement and onboarding Blind spots Familiar gaps can go unnoticed Less context on internal workarounds
Large enterprises often run a hybrid model. Internal teams self-assess each quarter, and an outside reviewer validates the critical domains once a year.
Either way, the same people take part, from the sponsor and governance lead to risk, compliance, legal, security, data owners and procurement. The business owner of each sampled system joins too. If the goal is formal certification, the ISO 42001 certification path runs through an accredited body, not an internal review.
AI Assessment
Take the AI Maturity Assessment
Score your AI foundations, generative AI and AI agent readiness in a few minutes and get recommendations for the gaps your self-assessment should look at first.
Start Your AI Assessment → How Long an Assessment Takes and What the Output Pack Contains
Duration depends mostly on the state of the inventory. A team with a clean, owned AI register can move straight to evidence, while one starting from spreadsheets spends the first weeks finding systems. Slow evidence requests and complex vendor estates stretch the timeline further.
For example, a sample four-week plan for one business unit with a usable inventory looks like this.
In week 1 , agree on the charter, confirm owners and reconcile the inventory.During week 2 , collect evidence, run interviews and pull samples.Then, in week 3 , test controls, complete the worksheet and score each domain.Finally, in week 4 , write findings, validate them with owners and agree on the backlog.At the end, the sponsor receives a short output pack with five parts. These are an executive summary, the worksheet with its evidence register, the tiered inventory, the gap register with its backlog, and a framework mapping. The mapping links results to NIST AI RMF or ISO 42001 and states the review’s limits. A tool such as Microsoft Purview can hold parts of that evidence trail where data governance and AI governance overlap.
Five Mistakes That Weaken an AI Governance Assessment
The same errors show up across industries, and each one makes the scores look better than the controls really are.
Assessing only the systems IT knows about. Vendor features and shadow AI stay invisible, so the riskiest systems never get scored.Accepting a policy as proof. A document shows design, while only records show operation.Scoring without evidence links. A score nobody can trace back to a record will not survive an auditor’s first question.Treating every gap as equal. Ten minor documentation gaps matter less than one agent with unrestricted write access.Closing findings on paperwork. An updated procedure closes nothing until a fresh sample shows it running.The fifth mistake is a common reason the same findings return a year later. It is also the easiest to prevent, because the closure test lives in the backlog item from day one.
When to Repeat the Assessment
Set the schedule by risk. High-risk systems and agents with write access deserve a review at least twice a year, while low-risk internal tools can wait for the annual cycle. Repeat the full assessment after a reorganization, a merger, or a new regulation that changes your obligations.
Trigger a targeted review whenever something material changes, such as a new model version, a new data source or new agent permissions. A vendor switching on AI features in a product you already use counts too. Between assessments, monitoring keeps the picture current, and our guides to AI observability tools , AI guardrails and AI agent lifecycle management cover the runtime side.
Finally, track two numbers from one cycle to the next. The first is control coverage, the share of in-scope systems with evidence for every critical control. The second is remediation closure, the share of backlog items closed on evidence by their due date. Together they show whether governance is improving or just being re-measured.
How Kanerika Runs AI Governance Assessments
Kanerika’s AI governance consulting team runs assessments in the order this guide describes, starting with the highest-risk systems. Common gaps include agents that inherited a human user’s permissions, vendor contracts silent on model training with client data, and responsible AI principles that never reached an approval workflow. The work moves through five stages.
Scope and discover. We agree on the charter, then build the inventory from interviews, SSO and procurement data, and agent registries.Collect and test. We sample evidence for the highest-risk systems first, then widen coverage.Score and report. We complete the six-domain worksheet with the critical-control override and write evidence-linked findings.Plan the fixes. We turn findings into a 30/60/90 backlog with owners and closure tests.Build and re-test. Where clients want help closing gaps, we build the controls and re-score them on fresh evidence.In addition, assessments often reveal that evidence lives in too many places. That is why our data governance services and AI governance work usually run together, since lineage, access policies and audit trails serve both. We also map findings to AI compliance obligations and the client’s existing AI GRC operating model , rather than creating a parallel process.
For a global expert-network firm, Kanerika built an AI compliance agent that vets experts against a compliance rulebook and writes structured reports with citations. Expert vetting became 3x faster, backlog cases fell by 70% and event delays dropped by 40%. Every decision now follows defined rules and leaves an auditable record, which is the kind of evidence an assessment looks for.
Case Study
3x Faster Expert Vetting with an AI Compliance Agent
Kanerika built an AI compliance agent that vets experts against a compliance rulebook with cited, auditable reports, cutting backlog cases by 70% and event delays by 40%.
Read the Case Study → Turn Your AI Policy Into Evidence
An AI governance assessment answers one question, whether your AI controls work and how you know. The answer lives in the inventory, the evidence register and the backlog, not in the policy binder. Run the seven steps, score what you can prove, and fix the critical gaps first.
Then run it again on a schedule, so that next Monday’s request from the audit committee takes an afternoon instead of a month. For a broader view of security controls that feed the same evidence, read our LLM security guide .
Frequently Asked Questions
What is an AI governance assessment?
An AI governance assessment is a structured review of whether an organization’s AI controls exist, operate and leave evidence. It inventories AI systems, collects records for each control, scores them on a worksheet, and turns every gap into a remediation item with an owner, a deadline and a clear test for closing it.
How do you conduct an AI governance assessment?
Run seven steps in order. Set the scope and owners, build the AI inventory, and collect evidence that controls operate. Then score a worksheet across six domains, write evidence-based findings, and convert them into a prioritized 30, 60 and 90 day remediation backlog. Repeat the cycle on a risk-based schedule afterward.
What questions should an AI governance assessment checklist include?
Cover six domains, namely accountability and policy, AI inventory and risk tiering, lifecycle controls, monitoring and evidence, third-party AI, and agent oversight. Each question should ask whether a control is demonstrated, such as whether every agent has an owner and documented permissions, and name the evidence an assessor must see.
What evidence is needed for an AI governance assessment?
You need both policy evidence and operational evidence. Policies, role charts and procedures show that a control is designed. Approval tickets, review minutes, monitoring alerts, incident logs, vendor reviews and agent action logs show it actually ran. Missing evidence should score zero rather than being treated as not applicable.
Who should be responsible for an AI governance assessment?
An executive sponsor owns the outcome, and a governance lead runs the assessment day to day. Risk, compliance, legal, security, data owners and procurement supply evidence, along with the business owner of each sampled system. Control owners can correct facts, but they should not score their own controls.
How long does an AI governance assessment take?
Duration depends mostly on the state of the AI inventory and how quickly owners return evidence. A single business unit with a usable inventory can follow a four-week plan, covering scoping, evidence, scoring and findings. Enterprises without an inventory spend extra time discovering systems before scoring can begin.
Can an organization perform its own AI governance assessment?
Yes. Self-assessment works well for a first baseline and for quarterly checks, because internal teams know the systems. Independence is its weak point, so many enterprises add an outside reviewer once a year for critical domains, board reporting or regulated systems where findings must carry weight.
Does an AI governance assessment establish ISO/IEC 42001 compliance?
No. An internal assessment can show how close you are to ISO/IEC 42001 and which gaps remain, but certification comes only from an audit by an accredited certification body. Treat the assessment as readiness evidence that prepares you for that audit, not as a substitute for it.
How is an AI governance assessment different from an AI audit?
An assessment is usually internal and diagnostic, aimed at finding gaps and planning fixes. An audit tests a system or program against a defined standard and produces a formal opinion, often from an independent party. Running an assessment first makes the later audit faster and far less surprising.
What are the metrics for measuring AI governance?
Two metrics travel well between assessment cycles. Control coverage is the share of in-scope systems with evidence for every critical control. Remediation closure is the share of backlog items closed on evidence by their due date. Add inventory completeness and incident counts by risk tier for a fuller view.
How often should you repeat an AI governance assessment?
Set the cadence by risk. High-risk systems and agents with write access deserve review at least twice a year, while low-risk tools fit an annual cycle. Run a targeted review after any material change, such as a new model version, new agent permissions or a vendor adding AI features.
