An AI agent is a capable actor placed at a seam — allowed to propose far more than it should be allowed to permit. Every HOT decision is really a decision about that line: who has the judgment to move it, who has the authority to set it, and whether the platform can hold it.
Both are capable-but-untrusted actors placed at a seam: they can do more than the institution yet trusts them to decide alone. You don't solve that by refusing the capability, and you don't solve it by handing over the keys. You solve it by building the muscle that lets a capable actor operate under supervision that shrinks over time as trust is earned.
That muscle is institutional, not personal. The ability to apprentice a junior person and the ability to govern an agent are the same capability wearing two costumes. An organization that has one tends to have the other; an organization that has neither will fail at both — and AI just makes the failure faster and more expensive.
Run the agent in shadow mode: it proposes, a human decides, and you count how often the human overrides. A real transformation is one where that override rate falls over time — the agent's proposals get better, the humans correct less, and the permit can safely widen because it earned the trust, not because someone got tired of saying no.
A flat override rate is the tell. It means the loop isn't learning — the capability isn't compounding, and no amount of additional model spend will change that. The falling curve is the signal that all three dimensions are actually wound together.
The falling curve is the whole game. It means judgment (H), authority (O), and enforcement (T) are compounding together — the only evidence that you built a capability rather than bought a demo.
The seam is not a technology decision, or an org-chart decision, or a training decision. It is all three at once — which is exactly why moving one dimension and calling it done never holds.