Agent Governance: A Practical Guide for Enterprise Leaders

Agent governance fails in the enterprise for one reason: it treats every agent the same. The firms getting this right are not the ones with the most policy. They are the ones with an inventory, an owner per agent, and a reversibility test that tells them which decisions actually need a human in the loop. This guide collects everything I have written on the subject from inside a consulting firm helping enterprises move AI from pilot to production, governed.

Start with the inventory you don't have

You cannot govern what you haven't counted. In client work we consistently find the real agent count is a multiple of the official one, and a meaningful share has no owner and no kill switch. You can't govern an inventory you don't know you have, shadow AI is shadow IT at 10x, and nobody owns the agent is the most common finding in the field. The first governance project is a census.

Govern by reversibility, not by policy volume

The governing criterion is simple: can the action be undone? Reversible actions need light-touch governance. Irreversible actions need a human. That single test sorts your entire agent estate into tiers, and it is why reversibility is the governing criterion and why uniform governance is how agents die, one-size governance either strangles the safe agents or waves through the dangerous ones. The analyst world is converging here too, but the practical version is what clients can actually run: the agent governance stack is finally here.

Keys, vendors, and the incident rate

Your agents hold credentials. Somebody changed the locks on the building; nobody changed them on the agents: your agents have keys. Vendor claims of governance features are mostly checkbox theater: your vendor checked the box, and the market is pricing the gap: the agent security gap just got a valuation. Watch the incident rate the way you watch uptime. Standards are coming, which helps: NIST is wiring the agent standards layer.

Why governance doesn't scale (and what does)

Most governance programs collapse under their own coordination cost: why governance doesn't scale, governance isn't your problem, uniform governance is. And the perimeter keeps moving: the day your agent got its own email address is the day your governance scope grew a mailbox. That day has arrived.

What to do Monday

1. Count the agents. Ask every team for their official number, then go find the real one.

2. Put a name next to every agent. No owner, no deployment.

3. Sort the estate by reversibility. Human-in-the-loop for the irreversible tier only.

4. Write the kill criteria before you need them, and assign who pulls the switch.