The Three Questions That Are Your Whole AI Governance

The follow-up to the governance gap: what you actually do about it.

A couple of weeks ago I wrote that most companies don't have an AI problem — they have a governance gap. The piece traveled further than almost anything I've written, and the replies all circled the same question: okay, I believe you. So what do I actually do?

Fair. It's easy to name a gap and leave people staring into it. So here's the answer, and the good news is it's much smaller than the enterprise-governance industry wants you to believe.

If you run a company of, say, twenty to a few hundred people, you do not need a forty-page AI policy. You do not need an ISO 42001 certification or a governance committee with a charter. Those frameworks exist, and they're real, and they're built for organizations with thousands of employees and a regulator watching. For everyone else, they're a way to feel governed without being governed — a document nobody reads protecting against a risk nobody mapped.

Right-sized AI governance is simpler and harder at the same time. It's the discipline to answer three questions for every agent you put into your operation. Not once, in a policy. Continuously, for each one.

Question one: Who owns this agent's output?

A person. Not a team, not “the AI committee,” not “whoever set it up.” One human whose name is attached to what this agent produces, the way a manager's name is attached to their team's work.

This is the question companies skip most, because it feels bureaucratic. It isn't. Ownership is what turns an agent from a floating capability into something accountable. When the invoice goes out wrong, the customer email lands badly, or the report has a number nobody can source, the first question is always “who owns this?” If the answer is a shrug, you've found your real problem — and it isn't the AI.

Question two: What is it allowed to decide alone — and what does it escalate?

Every agent needs a boundary, and the boundary is a judgment call you make, not one you let the tool make by default. The cleanest way I've found to draw it is to sort the agent's work into three buckets. Mechanical: rules-based, low-stakes, fully automatable — let it run. Pattern-based: it can draft or recommend, but a human signs off before anything ships. Judgment: context-heavy, high-stakes, reputation-adjacent — the agent supports, a human decides.

Most teams have never done this exercise even for their people, which is exactly why the agent inherits the ambiguity. Draw the line explicitly, write it down, and — this is the part that matters — make sure the agent actually escalates when it hits the edge, and that a specific human knows the escalation is coming. An escalation path nobody's watching is just a slower failure.

Question three: When it's wrong, what happens — and who catches it?

Not if. When. The most dangerous agent isn't the one that fails loudly; it's the one that fails quietly, inside a workflow nobody's looking at, producing confident output that's subtly off for three weeks before anyone notices.

So you need two things: a trail and a watcher. The trail is a simple record of what the agent did and on whose authority — Singapore's new agentic-AI framework actually requires every agent to carry a verifiable identity and an audit log of who authorized its actions, and while you don't need their machinery, you need their instinct. The watcher is the human from question one, checking the seams, especially where one agent's output becomes another's input. That handoff, unread, is where the real risk lives.

That's it. Three questions, answered honestly, per agent, kept current. You can capture it in a shared doc — a row per agent, three columns: owner, authority, escalation-and-audit. That single table is more governance than 71% of companies deploying agents right now actually have.

Here's the reframe I want to leave you with. Governance has a branding problem: everyone hears “the thing that slows AI down.” It's the opposite. Governance is the thing that lets you speed up — because the difference between an agent you can lean on and one you're quietly afraid of isn't the model. It's whether you can answer these three questions. Answer them, and you can hand the agent more, not less. That's not caution. That's how you actually get the return everyone was promised.