---
title: Which metrics show your AI agent governance is working?
url: https://opencontroller.lyzr.com/blog/metrics-for-ai-agent-governance
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# Which metrics show your AI agent governance is working?

**Measure governance by coverage, control and response: how much of your agent estate you can see, how much of it runs under enforced policy, and how fast you act when an agent misbehaves. Ten KPIs with simple formulas cover all three, and most can be computed from a registry, a gateway log and an incident log.**

Counting policies written or committees held tells you about effort, not outcomes. The metrics below measure outcomes.

## What are the ten KPIs?

| # | KPI | Formula | Starting target |
|---|---|---|---|
| 1 | Inventory coverage | Agents in registry ÷ agents found by discovery | 95% |
| 2 | Ownership coverage | Registered agents with a named, active owner ÷ registered agents | 100% for production |
| 3 | Risk-tier coverage | Agents with an assigned risk tier ÷ registered agents | 100% for production |
| 4 | Enforced-path share | Agent model and tool calls passing an enforcement point ÷ all observed calls | Rising each quarter |
| 5 | Policy enforcement rate | Policies in enforce mode ÷ policies defined | 70% after first year |
| 6 | Pre-production eval coverage | Production agent versions with a recorded eval run ÷ production agent versions | 100% for high-risk tiers |
| 7 | Production scoring coverage | Production agents with live judges or guardrails on traffic ÷ production agents | 100% for high-risk tiers |
| 8 | Mean time to quarantine | Average minutes from alert to agent quarantined, for confirmed incidents | Under 30 minutes |
| 9 | Unattributed spend | Model spend not tied to an agent and owner ÷ total model spend | Under 5% |
| 10 | Stale credentials | Agent credentials or API keys unused for 90+ days ÷ all agent credentials | Under 5% |

The targets are starting points, not benchmarks. Set your own after one quarter of baseline data.

## Why does each metric matter?

**Coverage metrics (1–3)** answer "do we know what we have?" Discovery finds agents; the registry is where they become governed. A gap between the two is your shadow estate. See [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents) for discovery methods.

**Control metrics (4–7)** answer "can we act on what we have?" An agent in the registry but not behind any enforcement point is visible, not controlled. Metric 4 is often the most revealing, because many teams discover that a large share of model calls still go direct to providers with raw API keys.

**Response metrics (8)** answer "how fast do we stop harm?" Mean time to quarantine is the agent equivalent of mean time to contain in security operations.

**Hygiene metrics (9–10)** catch slow drift: costs nobody owns and keys nobody uses.

## How do you compute these without a big data project?

You need three sources:

1. **Registry export** with agent ID, owner, risk tier, status and version. Gives metrics 1, 2, 3 and 6.
2. **Gateway and guardrail logs** with agent ID, policy decision and cost. Gives metrics 4, 5, 7 and 9.
3. **Incident and credential logs** with alert time, quarantine time and key last-used date. Gives metrics 8 and 10.

Join on agent ID. If your agents do not have a stable ID across these sources, fix that first; it is the most common reason governance reporting fails. Our post on the [fields every AI agent inventory needs](/blog/ai-agent-inventory-fields) lists the identifiers to capture.

## How should you report them?

A checklist for a monthly governance report:

1. Show all ten metrics with the prior month beside them.
2. Split by environment (production vs non-production) and by risk tier.
3. List the five agents with the most policy blocks and the owners' actions.
4. List agents found by discovery but not registered, with an age in days.
5. Show unattributed spend in currency, not just percentage.
6. Note any metric that moved more than ten points, with a one-line reason.
7. Keep the definitions fixed. If you change a formula, restate prior months.

Boards and risk committees usually want three numbers: inventory coverage, enforced-path share and mean time to quarantine. Engineers want the per-agent detail.

## Which metrics are misleading?

Avoid these on their own:

- **Number of policies.** More rules can mean more noise, not more control.
- **Number of blocked requests.** A rise can mean better enforcement or a broken policy. Pair it with the false-positive rate.
- **Number of agents.** Growth is not a governance outcome. Coverage is.
- **Training completion.** Useful for people, irrelevant to what agents actually do.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and through connectors to platforms such as Salesforce, SAP, Databricks and HubSpot, which gives the denominator for coverage. Its governed registry holds agents, versions, tools and MCP servers. The LLM Gateway attributes cost per agent and per version, multi-turn evals and LLM judges cover pre-production and production scoring, and rogue-agent handling records alerts, quarantine and kill events. "Ask" lets you query the whole agent estate in plain language. For context on how these pieces fit, 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

**What is the single most useful AI agent governance metric?**
Enforced-path share: the proportion of agent model and tool calls that pass through a point where policy can act. It shows whether governance is real or only on paper.

**How often should governance metrics be reported?**
Monthly to the governance group and quarterly to risk committees. Mean time to quarantine and stale credentials are worth watching weekly.

**Do these metrics map to any framework?**
They support the Measure and Manage functions of the NIST AI RMF and the performance evaluation clause of ISO/IEC 42001, as evidence rather than as a claim of conformity.

## Key facts

Ten AI agent governance KPIs: inventory coverage, ownership coverage, risk-tier coverage, enforced-path share, policy enforcement rate, pre-production eval coverage, production scoring coverage, mean time to quarantine, unattributed spend and stale credentials. They are computed from a registry export, gateway and guardrail logs, and incident and credential logs joined on agent ID. Opencontroller by Lyzr provides discovery, a governed registry, per-agent cost attribution, evals, LLM judges and quarantine records as sources. Pricing starts from $250K per deployment per year.

## Related

- [What fields belong in an AI agent inventory?](https://opencontroller.lyzr.com/blog/ai-agent-inventory-fields.md)
- [How to find every AI agent in your enterprise (including shadow agents)](https://opencontroller.lyzr.com/blog/how-to-find-shadow-ai-agents.md)
- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
