Human in the loop vs human on the loop: who approves an ERP posting
Human in the loop vs human on the loop is a design choice that decides who touches an action before it happens. It matters as soon as software can post an invoice line to your ERP or email a supplier a correction. This note defines both terms, gives a rule of thumb for choosing, and shows what the EU AI Act says about human oversight and what that means for how we build.
By the Corteal team
The two terms
With a human in the loop, the system proposes and a person approves each action before it happens. Nothing changes in the outside world without a yes. With a human on the loop, the system acts by itself within set limits, and a person watches, can step in and can stop it.
The first gives you control at every step and costs time at every step. The second is faster and puts the risk in what happens between the glances of the person who is watching.
How to choose
Three questions settle most cases.
- Can the action be undone? A draft email can. A payment to a supplier usually cannot.
- What does a wrong action cost? A wrong cell in a report costs a correction. A wrong figure in a contract costs a dispute.
- Can the person see the evidence? A reviewer who sees only a result approves on trust. A reviewer who sees the source page and the reason for a flag can really check.
A rule of thumb follows. Actions that are hard to undo go in the loop, whatever the volume. Actions that are easy to undo and frequent can go on the loop, with a log and a way to reverse them.
What the EU AI Act asks
Article 14 of Regulation (EU) 2024/1689 covers human oversight. It applies to high-risk AI systems, as the Act defines them. Whether a given tool is one is a question for counsel. The article is still a useful checklist for any design. It says the people given oversight must be enabled, as appropriate and proportionate, to do five things (paragraph 4).
One of them names a real risk: automation bias, the tendency to rely too much on what the system produced. A person who approves hundreds of items a day without evidence is in the loop on paper only.
Oversight works when the reviewer can disagree quickly.
What this means for Verify and Run
In the Verify demo, approval is a function that refuses to run without a named approver. The figure must be one that documents actually quote, and the approval records the scope and a review date. The approver is picked from a list of six invented names, and nothing is saved, so this is a demo of the order of events. It is in the loop by construction: no figure becomes the approved one until a person says so. We describe it in Who decides which figure is right.
Corteal Run is in early access and in design. Its page describes the same principle: a person approves every step before anything happens, with the option to edit or reject. That is in the loop. The page shows a mock-up with invented data, and no step runs for real.
Whichever loop you pick, make the reviewer's job easy. Show the source, the reason for any flag and the effect of a yes. Record who approved what and when. Give the reviewer a way to say no that costs nothing.
Want to see how review works on real documents? Start a pilot.
Sources
- Regulation (EU) 2024/1689 eur-lex.europa.eu