AI Governance Maturity Model: The 5 Levels Explained

Superblocks Team
+2

Multiple authors

October 20, 2026

0 min read

Copied
0:00

An AI governance maturity model is a five-level scale that shows how developed your AI oversight is, from ad hoc to optimized, and what to fix next.

Deloitte's 2026 State of AI in the Enterprise report found only 21% of organizations have a mature governance model for AI agents, so about four in five do not have one yet. This guide covers the five levels, a template for scoring yourself, and the controls that move you from one level to the next.

What is an AI governance maturity model? The 30-second answer

An AI governance maturity model is a staged framework that rates how consistently and provably an organization governs its AI. This guide scores it on five levels across four dimensions (data, process, people, and tooling).

Bottom line: A maturity model gives you one level you can defend and names the weakest dimension holding you at that level.

Key features

  • Levels: Five named levels, each defined by the evidence you can show.
  • Dimensions: Four areas scored separately, so one strong area can't hide a weak one.
  • Anchors: NIST AI RMF and ISO/IEC 42001, so scores map to standards auditors recognize.
  • Output: A current level, a target level, and a short list of controls to add.

Our AI governance guide covers the program a maturity model measures.

Why AI governance maturity matters

IBM's 2025 Cost of a Data Breach report, conducted by the Ponemon Institute, studied breaches at 600 organizations between March 2024 and February 2025. Its AI findings point to governance gaps.

What IBM found:

  • Breaches reach AI systems. 13% of organizations reported a breach of an AI model or application, and another 8% didn't know whether they'd had one.
  • No access controls. 97% of the organizations with an AI breach had no AI access controls.
  • No policy. 63% of breached organizations had no AI governance policy or were still developing one.
  • Higher breach cost. Organizations with high levels of shadow AI averaged $670,000 more in breach costs than those with low or no shadow AI.

A maturity level turns that exposure into a score you can track, showing whether the missing piece is inventory, ownership, enforcement, or measurement.

The 5 levels of AI governance maturity

Names vary by model. Databricks calls its levels Ad Hoc, Developing, Defined, Managed, and Optimized, and some models label Level 3 “Operationalized.”

This page uses one set of names throughout, matching the scale in our AI governance assessment, where Defined is 2 and Enforced is 3.

Level 1: Ad hoc

AI shows up wherever a team adopts it. There is no inventory, no named owner, and no review before something goes live. Shadow AI is the default state.

Observable signal: You can't list every AI system in use, and nobody is accountable for any of them.

Level 2: Defined

A written AI policy exists, an inventory has started, and key systems have named owners. Risk tiers are drafted. Only some systems are covered, and nobody can show evidence the policy is followed.

Observable signal: A policy and a registry exist, but only part of your AI is in the registry.

Level 3: Enforced (also called Operationalized)

Controls apply to every system, with records to prove it. Approvals gate launches, access runs through SSO with least-privilege roles, and test results are stored against each inventory entry.

Observable signal: For any AI system, you can hand an auditor the approval, risk rating, and test results today.

Level 4: Measured

Governance now has metrics. You monitor drift, exceptions, and audit findings, quantify the risk left after controls, and report KPIs to leadership.

Observable signal: A recurring governance report exists, and it has changed at least one decision.

Level 5: Optimized

Controls run inside the build and deploy pipeline, discovery is continuous, and incidents trigger policy updates through a documented process. Routine checks run without a person approving them.

Observable signal: A routine new AI use case gets a risk rating and controls without a manual review ticket.

AI governance maturity models compared

Several groups publish their own maturity models. They differ in how many levels they use and what they score, so pick one set of level names and stay with it.

How the main models compare:

🔍 Model 🪜 Levels 📊 What it scores 📝 Worth noting
Databricks 5: Ad Hoc, Developing, Defined, Managed, Optimized Five dimensions, including strategy and leadership, risk management, and monitoring Gives a 30-day baseline, a 90-day pilot, and 12-month and 24-month targets
IEEE-USA Not listed on the report page A questionnaire and scoring guidelines built on the NIST AI RMF Adjusts questions by lifecycle stage and by how many systems you assess
GenAI Global 4: Nascent, Emerging, Established, Leading Five domains, each with its own control considerations Lets you enter results online and compare against peer organizations
MITRE 5: Initial, Adopted, Defined, Managed, Optimized Six pillars and 20 dimensions Covers AI adoption broadly, with responsible use as one of the six pillars
Dataversity 5: Ad Hoc, Defined, Operationalized, Measured, Adaptive Data, process, and people Suggests metrics such as a Data Integrity Index and audit readiness
Scytale 5: Ad hoc, Developing, Defined, Managed, Optimized Four dimensions (policy and ethics, risk management, data governance, and monitoring and accountability) Includes a five-step assessment and a 90-day starter plan
WitnessAI 4: Foundational, Emerging, Operational, Embedded Eight dimensions, split into structural and operational, mapped to NIST AI RMF and ISO/IEC 42001 A vendor blog that treats agent governance as a separate surface

You can map any of these models onto this page's five levels, because every model above with published levels moves from informal practice to continuously improved practice.

The 4 dimensions to score: data, process, people, and tooling

Databricks defines maturity across three dimensions (data, process, and people) and scores five in its matrix. This page uses four, and tooling deserves its own dimension, because a policy with nothing enforcing it leaves a program at Level 2.

How each dimension breaks down:

  • Data: AI inventory, data sources per system, lineage, and rules for what staff can paste into AI tools.
  • Process: Risk classification, approvals, pre-release testing, and incident response.
  • People: Named owners, a cross-functional decision body, and training. A Center of Excellence can own this work, and our guide to creating a Center of Excellence walks through it.
  • Tooling: Registry, SSO and RBAC, audit logs, monitoring, and automated checks.

Score each dimension on its own. Your overall level is your lowest dimension, since a gap in one dimension limits the others.

If you've run our 28-question assessment, its seven domains roll up here. Inventory and data controls feed Data, risk classification and testing feed Process, and access, documentation, and monitoring feed Tooling.

The assessment covers People least, so score that dimension from your owner assignments and committee records.

WitnessAI splits dimensions into structural (policy and regulatory alignment, risk classification, data governance, and accountability) and operational (lifecycle management, transparency, monitoring, and incident response). It argues that programs can look strong on paper and weak in practice, which is why this page asks for evidence at every level.

Anchoring the model to NIST AI RMF and ISO/IEC 42001

A maturity model needs a published standard behind it, or it stays one team's opinion. Two published references give your levels a standard auditors recognize.

The two anchors:

  • NIST AI RMF 1.0: Released January 26, 2023, with four functions (Govern, Map, Measure, and Manage). It's voluntary and gives you the vocabulary for describing each control. IEEE-USA's AI Policy Committee built a flexible maturity model on it in July 2024, with a questionnaire and scoring guidelines.
  • ISO/IEC 42001: Published in December 2023, it sets requirements for an AI management system you can certify against. It gives auditors a standard to test.

In practice, use NIST's functions to name and group each control, and use ISO/IEC 42001 clauses as the evidence checklist once you aim for Level 3 and above.

Mapping the levels to the EU AI Act

The EU AI Act sets obligations for high-risk systems, and the evidence you collect at Level 3 lines up with them.

How the obligations map to evidence:

  • Risk management: Article 9 requires a documented risk management system that runs across the system's lifecycle. Your risk tiers and stored test results provide that evidence.
  • Data governance: Record data sources and lineage for each system, which you score under the Data dimension.
  • Technical documentation and record-keeping: Keep the registry entry with its approvals, plus the audit logs you score under Tooling.
  • Human oversight: Name an owner and a review body for each system, which you score under People.

The obligations stay as written, but the deadline moved. The Digital Omnibus pushed the high-risk deadline from 2 August 2026 to 2 December 2027, and penalties for the most serious violations still reach €35 million or 7% of global turnover. Treat Level 3 as the working target for any system in scope, and confirm scope with counsel.

How to find your AI governance maturity level in 5 steps

Scytale recommends basing the judgment on evidence (inventories, policies, ownership records, risk assessments, and monitoring logs). Run one check per kind of evidence before you fill in the template.

What each check looks for:

🔍 Check ❌ Lower maturity ✅ Higher maturity
1. Inventory Teams report AI tools when asked, and AI built into third-party SaaS goes uncounted A maintained list with an owner, purpose, data inputs, and vendor for each system, including AI inside SaaS
2. Policies Policies don't cover AI use, data handling, or employee responsibilities Documented approvals, prohibited uses, data restrictions, and an escalation path
3. Risk controls Every system gets the same review, or none Risk tiers decide how much review each system gets
4. Ownership Nobody can say who decides on an AI launch Named owners, a committee, and a RACI for AI decisions
5. Monitoring No checks after launch Recurring monitoring for drift and control effectiveness, with audit trails kept

Use the results to fill in the template below.

AI governance maturity model self-assessment template

Copy this table. For each dimension, mark the highest level where you can show the artifact today, then take the lowest result as your overall level.

The template:

📊 Dimension 1 Ad hoc 2 Defined 3 Enforced 4 Measured 5 Optimized ✍️ Your level
Data No AI inventory Inventory started, partial Complete inventory, data sources mapped Lineage tracked per system Inventory and lineage update automatically ___
Process No review before launch Written policy, risk tiers drafted Risk-based approvals gate launches Residual risk and exceptions measured Incidents feed policy updates ___
People No named owner Owners named for key systems Cross-functional body meets on schedule Leadership reviews governance KPIs Governance built into planning ___
Tooling Shared accounts, no logs Some SSO and logging SSO, RBAC, and audit logs on every system Drift and usage monitoring live Controls enforced automatically at build time ___

Three rules keep the score honest:

  • Evidence rule: A level counts only if you could hand an auditor the artifact today. If you can't, drop a level.
  • Floor rule: Your overall level is your lowest dimension. Fix that one first.
  • Re-score cadence: Run it annually, and again after any major AI launch.

The audit-trail test: Pick one AI system and ask your team for its complete audit trail. If they can't produce it fast enough to answer a regulator's request, WitnessAI's rule of thumb puts you at Level 1 or 2.

Scoring one AI system vs. your whole portfolio

IEEE-USA's model adjusts its questions to the system's lifecycle stage (planning, data collection and model building, or deployment) and to organizations assessing more than one system. You can apply the template above the same way, once per system and at each lifecycle stage.

How to score across systems:

  • Score high-risk systems individually. Run the template on each system in your top risk tier, because a portfolio average hides the system that would fail an audit.
  • Match evidence to lifecycle stage. A system still in planning can't show monitoring data, so score it on the controls that apply at that stage and re-score at deployment.
  • Roll up with the floor rule. Your program level is the lowest result across your in-scope systems.

How to move up one level

Each level adds a specific set of controls. These steps take you from one level to the next.

🪜 From Level 1 to Level 2: build the inventory

  • Run discovery across SSO logs, expense reports, and browser extensions to find AI nobody registered.
  • Create a registry with an owner, purpose, and data sources for every system.
  • Draft a use policy built on AI governance principles, and sort systems into three risk tiers.

🪜 From Level 2 to Level 3: enforce the policy

🪜 From Level 3 to Level 4: measure results

  • Define KPIs. Databricks lists model accuracy thresholds, drift detection rates, policy exception counts, audit-finding closure rates, and data integrity scores.
  • Quantify the risk that remains after controls. Databricks notes that Level 2 and Level 3 programs identify risk but don't track the exposure that remains.
  • Track lineage for every production model. Our AI model governance guide covers the lifecycle side.
  • Add AI risk to your enterprise risk register, and extend governance to third-party vendors and agents, as WitnessAI recommends.
  • Report to leadership every quarter.

🪜 From Level 4 to Level 5: automate controls

  • Run checks in your CI/CD pipeline so a failed control blocks a release.
  • Run discovery continuously.
  • Feed incidents and near-misses into policy and test updates through a documented process.
  • Reassess every year as AI capabilities and regulation change.

How long does each stage take?

No industry benchmark exists. Databricks publishes one vendor's timeline, and it's the most specific one I found.

Databricks' suggested timeline:

⏱️ Milestone 📅 Databricks’s suggested timing
Baseline assessment Within 30 days of launching a program
Pilot on the two or three highest-risk systems 90 days
Level 3 across all dimensions Within 12 months
Level 4 Over 24 months

Treat these as a guideline. Databricks calls Level 3 Defined and this page calls it Enforced, but the level numbers match. Real timing depends on how many AI systems you run and how many teams build them.

AI governance maturity metrics to track

Levels 4 and 5 depend on numbers. Two of the sources above publish metric lists you can borrow, and one metric below comes from this page.

The metrics:

📊 Metric 📝 What it measures 🔍 Source
Policy exception count Number of approved exceptions to a policy Databricks
Drift detection rate Share of model drift that monitoring detects Databricks
Audit-finding closure rate Share of audit findings closed Databricks
Data Integrity Index Percentage of models using certified datasets Dataversity
Explainability Ratio Share of AI decisions linked to metadata lineage Dataversity
Bias Remediation Time Average time to identify and correct bias Dataversity
Audit readiness Time needed to compile model documentation Dataversity
Registry coverage Share of AI systems in the registry with a named owner Our suggestion

Start with registry coverage and audit readiness. Both need only your registry and documentation, and both test the evidence rule directly.

Who should own AI governance maturity?

A level means little unless someone is accountable. Databricks says a policy should list each owner's name alongside their title, and recommends a RACI (Responsible, Accountable, Consulted, Informed) for each AI decision point. 

The decision points it names are model onboarding, risk assessment sign-off, data access approval, production authorization, and incident escalation.

Databricks also names who belongs on the cross-functional committee (AI practitioners, data governance, legal, compliance, and business leadership). Dataversity lists AI ethics boards, data steward councils, and model validation committees as common structures.

Start with two named people, an executive sponsor and a day-to-day owner. Programs stay at Level 2 when everyone shares governance, and nobody is accountable for the score.

A 90-day starter plan

Scytale outlines a 90-day plan built on four areas, matching the Level 1-to-Level 2 steps above.

The 90-day plan:

  • Inventory: List every AI system, including AI built into SaaS tools.
  • Policy: Publish an initial AI use policy.
  • Ownership: Name an owner for each system and a sponsor for the program.
  • Priorities: Pick the higher-risk systems to review first.

Common roadblocks that stall AI governance maturity

Scytale and WitnessAI name overlapping blockers. These four are the ones to plan around.

The blockers:

  • Shadow AI: Employees adopt tools faster than governance can review them, so the inventory falls behind.
  • Unclear ownership: When responsibility is split across legal, security, and data teams, reviews stall because no one owns them through completion.
  • One framework only: Governance built on a single regulation or internal policy misses requirements from other regulations and frameworks.
  • One-time effort: A governance review done once and set aside goes stale the moment a new AI system ships or an existing one changes.

AI governance maturity model vs. AI governance assessment: what's the difference?

A maturity model and an assessment answer different questions and run on different schedules.

The comparison:

🔍 Factor 🪜 Maturity model 📝 Assessment
Question How advanced is our program overall? Where are our control gaps right now?
Output One level (1 to 5) plus the weakest dimension Scored controls and a fix list
Cadence Annually Quarterly or before big launches
Evidence One artifact per dimension One artifact per control

The takeaway: Run the assessment to find gaps, then use its domain scores as evidence for your level. Our 28-question AI governance assessment supports that first step.

Does the model work for AI agents?

The model works for agents only if you score them separately. A level tells you whether controls exist and are applied, but agents act on their own and run on non-human identities that may sit outside your registry. 

Score agents as their own row. WitnessAI argues that mature governance for standard AI doesn't carry over to agents, which can sit at an earlier stage entirely.

Our shadow AI agents guide recommends assigning an owner to each agent, enforcing least privilege, and using short-lived credentials. The guide also covers finding unregistered agents by monitoring egress and reviewing OAuth grants.

Extra evidence for agents:

  • Owner: Every agent has a named owner and a stated purpose.
  • Access: Each agent runs with least-privilege permissions and short-lived credentials.
  • Discovery: Review egress traffic and OAuth grants to find agents missing from the registry.

What I liked and didn’t like about maturity models

Pros (what works)

  • ✅ One number for leadership. Leadership gets one number to react to, and a target level gives budget requests a clear basis.
  • ✅ Floor rule finds the weak spot. The lowest dimension sets your level, so the area nobody wanted to examine gets attention first.
  • ✅ Evidence doubles for audits. Because the levels map to NIST and ISO, the evidence you collect also supports an audit.

Cons (where it falls short)

  • ❌ Self-scoring rewards intentions. A team scoring itself can rate the program it planned, and the evidence rule checks the score against what runs today.
  • ❌ Agents can hide behind the average. One overall level can look strong while agent governance sits at Level 1.

Should you use a maturity model? My take

Use one when more than one team builds or buys AI, when leadership keeps asking “how are we doing,” or when customers ask for governance evidence.

A maturity model is a good fit if you:

  • Run AI in several departments with different owners.
  • Need a target level to justify headcount or tooling.
  • Face an audit, a customer questionnaire, or an EU AI Act deadline.

Skip the full model for now if you:

  • Run a single AI use case with one clear owner.
  • Have no inventory yet, so start with discovery and the assessment.

Best practices for using the model

These habits keep the model in use after the first scoring round.

The habits:

  • Fix the level names. Renaming levels between reports breaks comparison over time.
  • Score with two people. One from governance and one from engineering give you a check on inflated scores.
  • Set a target by risk. Low-risk systems can stop at a lower level than high-risk ones, and Databricks suggests Level 3 within 12 months as a first target.

Common mistake: Teams treat the level as a certificate. It's a snapshot, and it goes out of date when you ship a new AI system.

My verdict on AI governance maturity models

The level names matter less than the evidence rule. A level you can't back with artifacts is a guess, and scoring by your lowest dimension keeps the result realistic.

Level 3 matters most, because that is where governance becomes provable, and customers, auditors, and regulators ask to see proof. Levels 4 and 5 add value only after that foundation is in place.

Where Superblocks fits

A maturity model scores your whole program. Shadow AI is the new shadow IT, and ungoverned app-building is where it shows up first. Business teams can build internal apps with AI tools and no IT review, which leaves those apps out of the registry and drags the Tooling score down.

Superblocks is the governed enterprise vibe coding platform. Each app is registered as it's built; IT can list and administer apps through the Admin MCP, and apps run on SSO and role-based access with audit logs and a Security Agent review before publishing.

RBAC is available on Teams and up. Audit logs, SSO, VPC deployment, Admin MCP, and Security Agent reviews are Enterprise features.

Superblocks covers tooling for AI-built apps. Policy, named owners, and the review body are still your work.

At Flex, an app holding company IP ended up on a personal Replit account. After moving to Superblocks in its own AWS account, 16 of 18 departments were building 150+ apps within 30 days.

See how a governed internal app comes together on Superblocks, registry and RBAC included from the first build, with the Superblocks Quickstart Guide.

Or book a demo to see how Superblocks fits into your own path to a higher maturity level.

Frequently asked questions

What are the five levels of AI governance maturity?

The five levels of AI governance maturity are Ad hoc, Defined, Enforced, Measured, and Optimized. Each level adds evidence, from no inventory or named owner at Level 1 to automated, self-correcting controls at Level 5. Other models use different level names.

How long does each stage of AI governance maturity take?

Each stage of AI governance maturity takes roughly 30 days for a baseline assessment, 90 days for a pilot, 12 months to reach Level 3, and 24 months to reach Level 4, per Databricks' published timeline. That is one vendor's guideline, since no universal benchmark exists and timing depends on how many AI systems you run.

What is the difference between a maturity model and an AI governance assessment?

The difference between a maturity model and an AI governance assessment is scope. A maturity model rates how advanced your whole program is on a five-level scale, while an assessment scores specific controls at a point in time. Run the assessment first, then use its domain scores as evidence for your level.

Which framework should an AI governance maturity model be based on?

An AI governance maturity model should be based on NIST AI RMF and ISO/IEC 42001. NIST supplies the Govern, Map, Measure, and Manage vocabulary for controls, and ISO/IEC 42001 provides a certifiable management-system standard for evidence. IEEE-USA's published model builds on NIST.

What tool helps raise the tooling dimension for AI-built internal apps?

Superblocks raises the tooling dimension for AI-built internal apps by registering each app as it's built and applying RBAC automatically, with SSO and audit logs available on Enterprise plans. Dedicated governance platforms remain the right choice for purchased models and third-party AI.

Stay tuned for updates

Get the latest Superblocks news and internal tooling market insights.

You've successfully signed up

Request early access

Step 1 of 2

Request early access

Step 2 of 2

You’ve been added to the waitlist!

Book a demo to skip the waitlist

Thank you for your interest!

A member of our team will be in touch soon to schedule a demo.

8

production apps built

30

days to build them

10

semi-technical builders

0

traditional developers

8+

high-impact solutions shipped

2 days

training to get builders productive

0

SQL experience required

See full story →

See the full Virgin Voyages customer story, including the apps they built and how their teams use them.

Large cruise ship sailing in a harbor with a road lined with palm trees and cars in the foreground.
Why not Replit, Lovable, or Base44?

"Those tools are great for proof of concept. But they don't connect well to existing enterprise data sources, and they don't have the governance guardrails that IT requires for production use."

Superblocks Team
+2

Multiple authors

Oct 20, 2026