“Human in the loop” has become a phrase vendors add after the demo. A reviewer does not experience a phrase. They experience a case file. If the file cannot show what the agent saw and who was allowed to finish the work, the loop is theatre. Governed AI ops is our name for the version that survives that review.
Five fields, or it is not a trail
- Input — the document, record, or message the agent actually used.
- Draft — what the model proposed, kept rather than summarized away.
- Decision — close, hold, or escalate, and the rule or judgment behind it.
- Owner — a named person, or a named rule for the narrow class allowed to close without one.
- Time — so aging, the SLA window, and a later review share a clock.
A screenshot of a prompt is not this list. A confidence number is not a decision. “The system approved it” is not an owner. We build FinGuard-shaped packs so someone who was not in the thread can still reconstruct the close.
Where the human actually sits
The human sits on the exception classes written down before go-live, and on any close your policy says a person must sign. They should see the case while it is still inside the window, not after a weekly export. The agent may prepare the first pass. It does not become the accountable party.
There is a narrow case where a rule, not a person, may close: the class is named, the rule is named, and the trail still stores input, output, and the rule that permitted the close. Unowned autonomy — an agent finishing judgment calls because nobody configured a stop — is out of scope.
What we will not imply
We will not publish an accuracy rate for your book before we have seen it. We will not describe a chat log as an audit system. We will not staff a bench and call it AI labour. Governed labour means the model takes volume and a named owner remains on the exception.
If you also want that owner to be Xeres, with an SLA in the statement of work, the offer is Outcome BPO. If you want the trail designed before you choose who sits the desk, start at governed AI ops or write hello@xeres.tech.