What happens when the AI gets it wrong?
It gets caught by the thing that decided whether to trust it, not by somebody noticing. Every machine proposal passes a trust envelope that reads the agent's authority, the amount, and how often humans have overridden that agent before, and produces one of five outcomes: commit and tell somebody, wait for one approver, wait for two, escalate to a named person, or refuse. When a proposal turns out to be wrong, the human who overrode it is recorded, that override lowers the agent's measured agreement rate, and a lower agreement rate narrows what the agent may do next time without anybody adjusting a setting.
The question behind the question
Nobody is really asking whether the machine will be wrong. It will be. They are asking what the blast radius is, who finds out, and how long it takes.
Most answers to this are a variation on human in the loop, which is a phrase that means a person is shown something before it commits. That is necessary and it is nowhere near sufficient, because a person shown two hundred proposals a day stops reading them by the second week. The measurable version of this failure is an approver whose median time per decision is under ten seconds, and it is the single most common way an approval control stops working while continuing to exist.
What the envelope actually reads
An agent's own ceiling, which somebody granted it deliberately and which takes two people to widen. The value of the action against a limit set for that agent. The confidence of the determination. Whether a policy in force actually covers the action. Whether the action can be reversed. How many downstream records it touches. And the agent's own track record, which is arithmetic on what humans did with its previous output rather than a claim it makes about itself.
The most restrictive of those wins, and every one that applied is stored beside the outcome. So the answer to why was this queued is not a verdict. It is a list: this agent is at prepare-only, the policy it applied declares suggest-only, and the confidence was 0.62 against a threshold of 0.80.
The rule that does not bend
No agent holds an approval capability, at any autonomy level, and the authorization engine refuses one before it consults any grant. So a misconfigured role cannot produce a machine approval, and neither can a mistake in the envelope.
That is deliberately blunt rather than clever. Maker-checker with two machine parties has one accountable party, which is none.
What being wrong costs, concretely
A determination that a human overrides is recorded with what the machine proposed, what the person did instead, and why. That override is counted, and an agent humans keep overriding drops out of the healthy band and is capped at queue-for-approval automatically.
It also does not ask again. A finding somebody has ruled on is not re-raised, which matters more than it sounds: an agent that keeps surfacing a question you already answered is how an exception queue becomes a list people have learned to skip.
What this does not do
None of this makes a wrong determination right. It bounds what a wrong one can do on its own, records what it was derived from, and makes the agent narrower afterwards. If your objection is that a model should not be near the ledger at all, the honest answer is that no language model is called anywhere in this product: every determination is a rule, a statistic or a retrieval over your own history.