Human in the Loop

What can be undone and what can't

Most of what gets handed to AI can be undone. Have it draft something and if it's off, ask again. Have it sort files and if something's misfiled, move it. Put a checkpoint in front of steps like these and all it does is slow things down for nothing gained.

What can't be undone is a different kind of thing: sending an email, moving money, deleting a file, publishing something publicly, confirming a booking. As tools that carry out several steps on their own have become more common, these points have naturally multiplied.

So the tools coming out now split the flow into two kinds of stretches: ones that run on their own, and ones that stop and ask a person. Where that threshold gets placed is close to the entire personality of a given tool.

A person's job splits three ways

The first is blocking. Just judge whether something is okay to go out as is, and answer yes or no. It's the most common and the fastest, but if there's nothing on screen to actually judge by, the answer just becomes a reflexive yes.

The second is choosing. AI puts forward two or three options and a person picks one. The differences are visible, so a real judgment actually happens. The trade-off is that generating the options takes more time.

The third is fixing. A person touches up the result before it goes out. This takes the most effort, but it leaves something behind: the parts a person corrected can feed the next round of training or the next rule change.

When a checkbox is all that's left

People lean toward not doubting whatever a machine hands them. If the last ninety-nine were right, the hundredth probably will be too. Most cases where something goes wrong despite a person being stationed there trace back to exactly this.

Add volume on top and the check collapses entirely. Hundreds of items a day, most of them fine, and all that's left is the motion of flicking to the next screen. It gets worse when the reasoning isn't shown alongside the result. A conclusion with no "why" gives a person nothing to grab onto.

A culture where saying no is awkward makes it worse too. If rejecting something means more work, an explanation owed, and everyone else left waiting, a person leans toward letting it through. Keeping a check real means making "no" cost nothing extra.

Making the check actually happen

First, cut the volume. Don't route everything to a person; route only what's genuinely uncertain. Filter by a rule: cases where the AI's own confidence score comes back low, cases that look different from usual, cases with a large amount attached.

Second, put the reasoning on the same screen. What it looked at, what evidence it leaned on, has to sit right there for a person to actually retrace the steps. Third, leave room to undo it. A short window to cancel after something goes out lowers the pressure on the check itself.