Measure whether the controls work, not whether they exist. Count how many agents act through your enforcement layer, how many hold current evidence, what share of actions are held or refused and why, whether approvers actually review, how fast you can tighten a limit, and whether a sample of decisions can be rebuilt from the record. Policies and committees are inputs; these are results.
Why do most governance measures miss?
Because they count activity. Number of policies approved, number of risk assessments completed, number of people trained. Each is useful, and none tells you whether an agent was stopped when it should have been.
AI agent governance differs from broader AI governance in exactly this way: agents act, so governance has to be measured at the action. A filter that blocks offensive text is easy to count and says little about whether the agent can refund the wrong customer; we explain the gap in guardrails are not governance.
Which measures matter?
| Measure | What it shows | Warning sign |
|---|---|---|
| Coverage | Share of known agents acting through the enforcement layer | Any agent acting outside it |
| Unregistered agents found | How well discovery works | Found by an incident rather than by discovery |
| Certificate currency | Agents and capabilities with evidence in date | Lapsed certificates still at high depth |
| Verdict mix | Allowed, held, refused, withheld, by agent and capability | Sudden shifts in either direction |
| Deciding bound | Which bound decides most holds and refusals | One bound doing all the work |
| Approval quality | Change and rejection rates, review time, seeded-item catches | Approvals in seconds, near-zero changes |
| Time to tighten | From decision to the limit applying | Hours, not seconds |
| Reconstruction rate | Share of sampled decisions fully rebuilt from the record | Any gaps |
| Incidents and near misses | Count, severity, link that failed | Repeats of the same cause |
| Change discipline | Bound changes with evidence and second approval | Changes with neither |
Ten measures is plenty. Pick the ones that match your largest risks and report them consistently.
What are the practical steps?
- Define the measures in your policy, so they are not invented after an incident. Our answer on how to write an AI agent governance policy covers where they fit.
- Pull them from the record, not from surveys. The verdict record, the register and the change log hold most of the data.
- Set thresholds that trigger review, such as any lapsed certificate at depth 3 or above.
- Test some directly. Time a limit change. Rebuild ten decisions. Seed test items into approval queues.
- Report monthly to owners, quarterly to the board, with trends rather than snapshots.
- Feed them into your control library. Many measures are control test results; our guide on connecting an AI agent to your governance, risk and compliance platform shows how to keep them in one place.
A worked example
Take a consumer lender's first quarterly governance report. Coverage looks strong: all seven known agents act through the enforcement layer. But discovery finds an eighth, built by the marketing team, that had sent offers to existing customers. Certificate currency shows one collections capability running at depth 3 on a lapsed certificate for two weeks, because the renewal run had been postponed. Time to tighten is measured at four hours, because limit changes needed a ticket to the platform team.
None of these would have appeared in a count of policies approved. The lender adds a discovery check to its monthly routine, makes lapsed certificates drop depth automatically, and gives owners a direct way to tighten limits.
Who owns the measures?
The governance lead compiles them. Process owners answer for their agents' numbers. Risk challenges. Internal audit tests a sample of the measures themselves, because a measure that is computed wrongly is worse than none.
Which controls keep the measures honest?
- Measures computed from the record, not self-reported.
- Definitions fixed in policy, changed only with approval.
- Independent testing by internal audit.
- Trends reported alongside thresholds, so slow drift is visible.
The topic guide, agent governance is a bound, not a model, sets out what the measures are measuring.
Questions people ask
Is a low refusal rate good?
Not on its own. It can mean the agent is well scoped, or that a bound was loosened. Look at which bound decides refusals and how that changes.
Should we benchmark against other companies?
There are no reliable shared figures for agent governance. Benchmark against your own trend and your own thresholds.
Who should see the measures?
Owners monthly, risk and the executive committee quarterly, and the board in summary. Each audience needs fewer measures than the one before.