---
title: What does an AI agent governance maturity model look like?
url: https://opencontroller.lyzr.com/blog/ai-agent-governance-maturity-model
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What does an AI agent governance maturity model look like?

**An AI agent governance maturity model has five levels, from ad hoc to continuously improving. Each level should be judged on observable numbers, not intentions: how much of the agent estate is inventoried, how often discovery runs, what share of agents have a named owner, whether policies are enforced at runtime, and whether evidence is produced automatically.**

Most maturity models fail because their levels are described in adjectives ("defined", "managed"). The version below ties every level to criteria an auditor or a board member can check.

## What are the five levels?

| Level | Name | Inventory coverage | Discovery | Agents with named owner | Runtime enforcement | Evidence |
|---|---|---|---|---|---|---|
| 1 | Ad hoc | Unknown | None | Unknown | None | None |
| 2 | Inventoried | Known agents only, in a spreadsheet | Manual, yearly or on request | Under half | None | Manual screenshots |
| 3 | Governed | Registry covers sanctioned platforms | Automated, at least weekly | Most production agents | Some: gateway routing and logging | Collected per audit |
| 4 | Controlled | Registry covers sanctioned and discovered agents, including SaaS and developer tools | Continuous or daily | All production agents | Policies block, redact or quarantine in the traffic path | Generated continuously |
| 5 | Improving | As level 4, plus agent dependencies (tools, MCP servers, models) | Continuous | All agents, with deputies | Enforcement tuned from incident and eval data | Mapped to frameworks and reviewed by management |

Set your own thresholds for "most" and "under half", but write them down. A maturity level you cannot measure will not survive a second assessment.

## How do you measure each criterion?

1. **Inventory coverage.** Divide agents in the registry by agents found by discovery. If you have no discovery, you cannot compute this, which is itself a level 1 or 2 signal.
2. **Discovery frequency.** Record when each source (cloud account, SaaS platform, developer device fleet) was last scanned. Report the oldest one, not the average.
3. **Ownership.** Count production agents with a named, current owner in the identity provider. An owner who has left the company does not count; see [what happens when an agent owner leaves](/blog/what-happens-to-ai-agents-when-the-owner-leaves).
4. **Runtime enforcement.** Count agents whose model or tool traffic passes through a control point that can allow, route, redact or block. Logging only is observation, not enforcement.
5. **Evidence automation.** Time how long it takes to answer "show me every agent that touches customer data, its owner and its last evaluation". Minutes suggests level 4; weeks suggests level 2.

These five numbers also work as a quarterly scorecard. For more metrics, see [metrics for AI agent governance](/blog/metrics-for-ai-agent-governance).

## How does the model line up with recognised frameworks?

The levels are not a certification. They are a way to show progress against frameworks you may already use:

- **NIST AI RMF.** The [AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) has four functions: Govern, Map, Measure and Manage. Levels 2 and 3 mostly address Map (knowing what you have). Levels 4 and 5 address Measure and Manage.
- **ISO/IEC 42001.** The [AI management system standard](https://www.iso.org/standard/42001) expects an organisation to plan, operate, monitor and continually improve its AI management system. Level 5's feedback loop is evidence for that improvement cycle.
- **EU AI Act.** For high-risk systems, logging and human oversight duties make level 4 enforcement and evidence relevant. Check the current application dates, as parts of the timeline have been subject to change.

Use the language "evidence for" these frameworks. A maturity level does not make an organisation compliant with any of them.

## What is a realistic path from level 1 to level 4?

A typical sequence:

1. **Run discovery first.** You cannot set ownership targets until you know the denominator.
2. **Assign owners** to everything in production, and retire what nobody claims.
3. **Put a control point in the traffic path** for the highest-risk agents, such as an LLM gateway or agent gateway.
4. **Write policies as code** and start them in monitor mode before blocking.
5. **Automate evidence**: export registry, policy decisions and evaluation results on a schedule.

Do not try to reach level 4 everywhere at once. Many teams run level 4 for high-risk agents and level 3 for internal low-risk ones, and say so openly in their assessment.

## How Opencontroller does this

Opencontroller covers several of these criteria directly. Auto-discovery finds agents, models and API keys across AWS, Azure and Google Cloud, and connectors reach third-party platforms (Salesforce, SAP, Databricks, HubSpot) and Claude Enterprise and OpenAI Enterprise workspaces. A governed registry records agents, versions, tools and MCP servers. The LLM Gateway and Agent Gateway put control in the path, and policy as code ships with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. Enforcement depth depends on the connector and traffic path: registration is not runtime takeover. Read [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**How often should we reassess AI agent governance maturity?**
Quarterly is common. The five numbers in this model can be produced each quarter, so the assessment becomes a report rather than a project.

**Can we be level 4 without runtime enforcement?**
No. Level 4 is defined by policies that act in the traffic path. Without enforcement, strong inventory and ownership put you at level 3.

**Is a maturity level the same as compliance?**
No. A maturity level is an internal measure. It can provide evidence for frameworks such as NIST AI RMF or ISO/IEC 42001, but it is not a certification.

## Key facts

- A five-level AI agent governance maturity model: ad hoc, inventoried, governed, controlled, improving.
- Each level is judged on five observable criteria: inventory coverage, discovery frequency, share of agents with a named owner, runtime enforcement and evidence automation.
- Inventory coverage is computed as agents in the registry divided by agents found by discovery.
- Level 4 requires policies that allow, route, redact, block or quarantine in the traffic path; logging alone is observation.
- The NIST AI RMF has four functions: Govern, Map, Measure and Manage.
- Maturity levels provide evidence for NIST AI RMF and ISO/IEC 42001; they are not certifications.
- Opencontroller provides auto-discovery, a governed registry, LLM and Agent Gateways, and policy as code with mappings to OWASP, MITRE ATLAS, NIST and ISO 42001.
- Opencontroller pricing starts from $250K per deployment per year.

## Related

- [Which metrics show your AI agent governance is working?](https://opencontroller.lyzr.com/blog/metrics-for-ai-agent-governance.md)
- [What is an AI agent registry?](https://opencontroller.lyzr.com/blog/what-is-an-ai-agent-registry.md)
- [Who should own AI agent governance: security, IT or the AI team?](https://opencontroller.lyzr.com/blog/who-owns-ai-agent-governance.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
