Human in the Loop: When Oversight Is Just an Automatic Click

The phrase that sounds responsible and almost never means anything

Every time someone presents an autonomous system, the same reassuring line appears: "there is always a human supervising". It sounds responsible. It almost never means anything.

Our CTO and AI lead, David Botella, put it plainly on LinkedIn. Putting a human in the loop is not a quality guarantee. It is an architecture decision. And choosing the wrong model for the wrong process is where money gets lost.

• A human in the process does not, by itself, equal effective control.
• The real risk appears when approval becomes a reflex, not a judgment.
• Oversight only counts if someone can say no with information, time and political room to do so.

Three loops, three different risks

In practice there is not one Human in the Loop. There are at least three models, and each one carries a different risk.

1. Human in the loop
   • The human sits inside the cycle, the system stops and does not move until someone decides.
   • It is the safest model and the most expensive in latency.
2. Human on the loop
   • The system runs on its own and the human watches, with real power to interrupt and reverse.
   • They do not approve every step; they intervene when the risk demands it.
3. Human out of the loop
   • The system acts and review comes later, through audits and traces.
   • It makes sense when the error is cheap and reversible.

Choosing the wrong model for the wrong process is exactly where money gets lost: either you slow down what should flow, or you let run what should not.

An Approve button is not judgment

This is where most teams get it wrong. Adding an "Approve" button is trivial. The hard part is designing a decision point where the human actually contributes judgment.

When someone reviews two hundred outputs a day, approval stops being a judgment and becomes a reflex. That is not oversight, it is an automatic click and legal responsibility stacked on someone who never had the information to say no.

Four signals for placing the checkpoint

David's criteria as CTO / AI Architect for placing those points are clear and actionable:

• Irreversibility. If the action cannot be undone (payments, shipments, deletions, external communication), the human enters before.
• Asymmetric cost of error. A slightly odd tone in a draft is not the same as a misread clause in a contract.
• Ambiguity of the rule. Where the rule is explicit, let the machine verify. Where intention must be interpreted, judgment enters.
• Ownership of responsibility. If someone with a name and surname is accountable, that person decides.

If it approves 100%, it protects nothing

There is something almost nobody builds. Measuring the checkpoint. If your reviewer has spent six months approving 100% of what arrives, that control protects nothing. It only creates latency and false safety.

The goal is not to put humans everywhere or remove them everywhere. It is to spend human attention where it changes the outcome.

Questions before you click the next Approve

Before treating a human-supervised flow as solid, it is worth asking:

• Which loop model are you actually using: in, on, or out of the loop?
• Does the reviewer have context, time and real freedom to say no?
• Are you measuring the checkpoint rejection rate, or only the latency?
• Are there approvals that have been an automatic click for months?
• If that control failed tomorrow, who owns the responsibility by name?

Spend attention where it changes the outcome

If you recognise in your flows an approval that is no longer a judgment but an automatic click, it is worth redesigning that point. Not to add more friction, but to put human judgment where it actually changes the result.

If you want to review together how supervision is set up in your AI systems, let's talk. A first map of checkpoints, with no commitment attached, usually makes clear where there is real control and where there is only a signature.

Read David Botella's original LinkedIn post