Human in the loop vs human on the loop: when should AI ask first?
Human in the loop means the AI asks a person before it acts. Human on the loop means the AI acts on its own, and a person checks afterwards and can step in. When you let AI do real work, the main decision is which tasks get which of the two.
An everyday comparison
Think of a new employee. In the first weeks they show you every letter before it goes out. That is human in the loop. Later they send routine letters themselves, and you look through a few at the end of the week. That is human on the loop. With AI the choice is the same.
| Human in the loop | Human on the loop | |
|---|---|---|
| The AI | Suggests | Acts |
| The person | Says yes or no first | Checks afterwards |
| Speed | As fast as the person | Immediate |
| Fits | Costly steps that are hard to undo | Frequent, small steps that are easy to undo |
Two questions that decide it
For anything an AI could do, ask:
- Can it be undone easily? A changed date can. An email to a customer cannot.
- What does a mistake cost? A wrong label costs a minute. A wrong payment costs money and trust.
If it is easy to undo and cheap to get wrong, let the AI act and check later. If it is hard to undo or expensive, the AI should ask first.
A starting list for a small company
- The AI may act alone: sorting and labelling, summaries, internal notes, reminders.
- The AI suggests, a person confirms: changing dates or who is responsible, changing many entries at once, replies to customers.
- Always a person: payments, contracts, deleting information, hiring and pay, deciding who may see what.
How to keep asking first from slowing things down
Asking first fails when the requests pile up and nobody looks at them. Three habits prevent that.
- Show the change itself. "Due date: from 3 March to 10 March" can be confirmed in a second.
- Put similar items together. Ten date changes in one question, not ten separate ones.
- Name a stand-in and a time limit. If nobody answers in two days, the question goes to someone who can.
As trust grows, you can loosen the rule. If you have confirmed the same kind of change many times without correcting it, let it run alone.
Keep a record either way
Whether the AI asks first or acts alone, keep a record of what it did and when. If someone approved it, note who. This record lets you check things later. It also shows you where it is safe to give the AI more freedom.
How Worknano handles this
Agents in Worknano suggest changes and wait for your yes before they change anything. You see exactly what will change and confirm it. For decisions that belong to a manager or to the finance team, Approvals add sign-offs by several people in a set order, with time limits and a permanent history. Every agent can do only what the person who asked is allowed to do. So it can never change more than that person could.
Create a free workspace and watch an agent ask before it acts.