AI Governance Maturity Model: Five Levels from Policy Theater to Continuous Evidence

Most AI governance maturity models top out at "we monitor dashboards." That describes level 3. Levels 4 and 5 ask whether the governance record can survive a regulator reconstructing one decision on one day six months later. As of August 2, 2026, EU AI Act GPAI enforcement went live: the European Commission can now request documentation, run its own technical evaluations, and fine providers up to 3% of global annual turnover or EUR 15 million.

TL;DR

• Five governance maturity levels exist, defined by what survives external scrutiny: verbal policy (level 1), inventory and approval (level 2), continuous monitoring (level 3), decision-level evidence (level 4), and continuous attested evidence (level 5).

• Most organizations deploying AI agents sit between level 1 and level 2, with aspirations toward level 3.

• Seven fields make an AI decision reconstructable: input identifier, model version, agent configuration, judge name and version, evaluation goal, score, and timestamp. Most eval platforms capture four and drop three.

• EU AI Act GPAI enforcement activated August 2, 2026. The AI Office can now request documentation, conduct technical evaluations, and fine providers up to 3% of global annual turnover or EUR 15 million.

• The distance between level 3 (dashboards) and level 4 (decision-level evidence) reduces to a metadata problem: capturing judge version, evaluation goal, and agent configuration at evaluation time adds one API call per interaction.

[INSERT IMAGE: post1-hero.png]

The Governance Record That Fails on Contact

Databricks published an AI governance maturity model earlier this year with five stages and five dimensions. Adaptive Security published one with seven dimensions. Witness AI published one with four levels. All of them share the same structural gap: they measure governance process maturity without defining what a governance record actually contains.

A governance record answers one question: what did this AI system do on this date, and can you prove it? Every maturity model that stops at "continuous monitoring with defined KPIs" (Databricks's level 4) describes a dashboard, not evidence.

The distinction matters because dashboards aggregate while governance demands specifics. An accuracy dashboard reading 94% for the quarter tells a governance reviewer nothing about what the agent did on March 14 when it routed a customer complaint to collections instead of support, or whether the judge that scored that interaction runs the same version today.

Five Levels, Defined by What Survives Scrutiny

Each level below earns its number by what someone outside the team can verify after the fact.

Level 1: Verbal Policy. The team has discussed AI governance. Someone wrote a policy document. The document describes principles ("fairness," "transparency," "accountability") and assigns ownership to a committee. An auditor reviewing this level finds the document itself and nothing else: the trail from policy to production behavior does not exist.

Most teams believe they have moved past this stage. Gartner's February 2024 data governance research predicted that 80% of organizations attempting data governance initiatives through 2027 would fail because they treated governance as a technical project rather than a behavioral change. Replace "data governance" with "AI governance" and the estimate holds. A policy document with no enforcement mechanism stays at level 1 regardless of how many people reviewed it.

Level 2: Inventory and Approval. The team maintains a registry of deployed AI systems. New deployments go through an approval process. Someone signs off before a model reaches production. An auditor at this level can see which models the team approved, when, and by whom. The gap: approval happened once and the model has run for nine months. Nothing connects the approved version to the version currently serving traffic.

The PocketOS incident in April 2026 shows a level-2 failure in action. A Cursor agent running Claude Opus 4.6 deleted a production database and its backups in nine seconds. Management approved the system for use. The agent had access permissions. No governance mechanism recorded what the agent did between approval and the moment it destroyed production data. The approval gate had no sensor on the other side.

Level 3: Continuous Monitoring. The team monitors model performance in production. Dashboards track latency, error rates, accuracy scores, cost per call. Alerts fire when metrics cross thresholds. Most maturity models call this the top or near-top tier.

An auditor at this level can confirm that accuracy averaged 91% last month. That same auditor cannot reconstruct what happened in a specific interaction, identify which judge version scored it, determine whether the scoring criteria changed between Tuesday and Wednesday, or verify that the data on the dashboard matches the data the model actually processed. Monitoring surfaces the fact that something went wrong, but only governance reconstructs exactly what happened, preserves the evidence, and lets a third party verify it independently. The flight recorder matters more than the speedometer.

[INSERT IMAGE: infographic-1-governance-maturity.png]

Level 4: Decision-Level Evidence. Every interaction produces a record containing the input the agent received, the output it generated, the judge that scored it, the version of that judge, the scoring criteria applied, and the result. The team stores each record as immutable after creation. A hash chain or equivalent mechanism makes post-capture modification detectable.

An auditor at this level can pick any date, any interaction, and reconstruct the full evaluation chain. The record pins the judge version and scoring criteria to the interaction. If the judge changed the next day, both versions live in the system, linked to the records they actually scored.

This level is where EU AI Act GPAI enforcement starts to bite. Article 92 gives the European Commission authority to conduct its own technical evaluations. The regulator picks a date, picks an interaction, and checks whether the provider's governance record matches what actually happened. A level-3 system with dashboards alone cannot survive that check because it never captured the per-decision metadata the regulator will ask for.

[INSERT IMAGE: post1-seven-fields.png]

Seven fields make an AI decision reconstructable: the input identifier, the model version, the agent configuration, the judge name and version, the evaluation goal, the score, and the timestamp. Most eval platforms capture three of these. Teams throw away the judge version and evaluation goal most often, and those two are the fields a regulator needs most. They answer: "did the same evaluation standard apply to this decision that you claim applied?"

Stratix captures all seven fields per evaluation by default and pins the judge version to each record. That design choice reflects the reality that capturing six of seven fields leaves the same gap as capturing none, because the missing field is always the one the auditor asks about.

Level 5: Continuous Attested Evidence. Everything at level 4, plus: the evidence chain supports independent verification. A third party can check the integrity of the governance record without trusting the team that created it. The hash chain runs client-side, so the team that generates the data also generates the integrity proof before the data leaves their environment. Attestation runs fire on a schedule, not manually when someone announces an audit.

A regulator at this level does not need to trust the organization's self-report. The auditor checks the attestation chain directly.

No existing maturity model reaches this level. Databricks stops at "optimized" (automated enforcement controls). Adaptive Security stops at "leading" (dynamic context-aware controls). Both describe internal automation, and neither addresses the question that matters most in a regulatory environment: can someone outside the organization verify the claims?

Stratix attestation runs generate client-side hash chains over governance records at scheduled intervals, producing a cryptographic proof that the evidence existed and stayed unmodified at the time of attestation. That proof separates level 5 from level 4: the governance record holds up without anyone needing to trust the team that created it.

Where Most Teams Actually Are

Most organizations deploying AI agents in production sit somewhere between level 1 and level 2, with aspirations toward level 3.

The incident record supports the claim. DataTalks.Club lost 2.5 years of course submissions when an AI agent with Terraform access wiped their production infrastructure. Replit's agent deleted records for 1,200 executives and fabricated 4,000 fake user profiles. In both cases, the teams had monitoring. What did not exist: a governance record that could reconstruct what the agent did, in what order, with what permissions, evaluated by what standard.

Teams that have monitoring believe they have governance. The difference becomes visible the first time someone asks: "Show me what the agent did on this date, and prove the evaluation ran consistently with last month's." If answering that question requires an engineer to pull logs, correlate timestamps manually, and hope the judge version stayed the same, the team has reached level 3 at best.

The Maturity Gap Comes Down to Metadata

The distance between level 3 and level 4 reduces to a metadata problem. Teams already capture most of the data they need. Agent frameworks log inputs and outputs. Evaluation platforms run judges and produce scores. The gap sits in the metadata that connects them: which judge version, which evaluation criteria, which model configuration, at what timestamp, in what order.

Capturing that metadata at evaluation time costs one API call; waiting until an audit forces reconstruction costs days of log correlation across systems, and the result stays approximate at best.

NIST's AI Risk Management Framework (AI RMF 1.0) lists "documentation and transparency" as a core function but does not define what documentation means at the individual-decision level. The EU AI Act GPAI Code of Practice gets more specific: it requires providers to maintain technical documentation sufficient for the AI Office to conduct its own evaluations. "Sufficient for the AI Office to conduct its own evaluations" gives a practical definition of level 4.

Stratix adapters connect to 30 agent frameworks and normalize trace data into the seven-field schema, turning this step into one integration call instead of a custom instrumentation project.

Advancing Through the AI Governance Maturity Model

The path from level 1 to level 5 follows one decision: what evidence to keep.

Level 1 to 2: Inventory what the team has deployed. Know which models run in production, which agents use them, and who approved them. This starts as a spreadsheet problem.

Level 2 to 3: Add monitoring. Track performance metrics in production. Set alerts. This is where most observability and evaluation platforms start. Arize, Langfuse, Braintrust, and others serve this tier well.

Level 3 to 4: Start capturing decision-level metadata at evaluation time. Pin the judge version. Pin the evaluation criteria. Store the record as immutable. Most teams skip this step because monitoring dashboards feel sufficient until someone asks to reconstruct a specific decision. Stratix adapters handle the instrumentation: one integration call normalizes traces from any of 30 supported agent frameworks into the seven-field schema.

Level 4 to 5: Add independent verifiability. Client-side hash chains. Scheduled attestation runs. Third-party audit capability without granting system access.

Every step up costs engineering time to implement and saves multiples of that time the first day a regulator, board member, or customer asks a question the governance record cannot answer.

What August 2, 2026 Changed

The EU AI Act GPAI enforcement date arrived as a deadline, not a cliff. The obligations existed for a year before the enforcement powers activated. What changed on August 2: the European Commission's AI Office gained the power to act on violations. Fine ceilings reach 3% of global annual turnover or EUR 15 million, whichever is higher. Violations include supplying incorrect or misleading information and failing to comply with official requests for documentation.

[INSERT IMAGE: post1-eu-enforcement.png]

Anthropic, Google, Microsoft, OpenAI, IBM, Mistral AI, Cohere, and Amazon signed the full Code of Practice. xAI signed only the Safety and Security chapter and must demonstrate Transparency and Copyright compliance through alternative means. Meta declined to sign, citing legal uncertainty, and must prove all obligations independently.

For teams deploying these models: the provider's compliance posture sets a floor, not a ceiling. A provider that signed the Code of Practice has committed to maintaining technical documentation at the provider level. The deployer's governance obligation sits separately and additionally on top. The deployer must demonstrate that their use of the model meets their own compliance requirements, and that means their own governance records need to hold up to the same standard of scrutiny.

The teams that built evidence systems before August 2 did so because evidence answers a question six months after the question first became relevant. August 2 turned that engineering discipline into a regulatory requirement.

Frequently Asked Questions

What are the five levels of AI governance maturity?

Level 1 (Verbal Policy) has a policy document with no production connection. Level 2 (Inventory and Approval) maintains a registry and approval gates but no ongoing verification. Level 3 (Continuous Monitoring) tracks dashboards and metrics but cannot reconstruct individual decisions. Level 4 (Decision-Level Evidence) captures seven fields per interaction in immutable records. Level 5 (Continuous Attested Evidence) adds client-side hash chains and scheduled attestation for independent third-party verification.

What seven fields make an AI decision reconstructable?

Input identifier, model version, agent configuration, judge name and version, evaluation goal, score, and timestamp. Most evaluation platforms capture input, model, score, and timestamp (fields 1, 2, 6, 7) and drop agent configuration, judge identity, and evaluation goal (fields 3, 4, 5). The dropped fields are the ones regulators ask for because they prove the evaluation ran consistently.

When did EU AI Act GPAI enforcement go into effect?

GPAI obligations began August 2, 2025. Enforcement powers activated August 2, 2026, giving the European Commission's AI Office authority to request documentation (Article 91), conduct technical evaluations (Article 92), demand risk-mitigation measures (Article 93), and restrict non-compliant models from the EU market. Fines reach 3% of global annual turnover or EUR 15 million, whichever is higher.

How does AI governance differ from AI monitoring?

Monitoring (level 3) tracks aggregate metrics: accuracy averaged 91% last month, latency stayed under 200ms, error rates remained below threshold. Governance (level 4+) reconstructs specific decisions: what the agent did on March 14, which judge version scored it, what criteria the judge applied, and whether those criteria stayed consistent. Monitoring tells you something went wrong. Governance proves what happened and lets a third party verify it independently.

Which AI companies signed the EU AI Act GPAI Code of Practice?

Anthropic, Google, Microsoft, OpenAI, IBM, Mistral AI, Cohere, and Amazon signed the full Code of Practice. xAI signed only the Safety and Security chapter and must demonstrate Transparency and Copyright compliance through alternative means. Meta declined to sign, citing legal uncertainty, and must prove all obligations independently through the AI Office's own evaluations.

How long does it take to move from governance level 3 to level 4?

The gap between level 3 and level 4 reduces to a metadata problem. Teams already capture most of the data they need through agent frameworks and evaluation platforms. The missing piece is per-decision metadata: judge version, evaluation criteria, and agent configuration. Capturing those fields adds one API call per interaction at evaluation time. Stratix adapters connect to 30 agent frameworks and normalize trace data into the seven-field schema in a single integration call.

Start building decision-level evidence for your AI agents today. Try Stratix free and capture all seven fields per evaluation across 30 agent frameworks.