When teams design their first AI automation, the conversation usually jumps straight to what the AI will do. The more important question is what a person will still do. Get that wrong in one direction and the automation saves nobody any time. Get it wrong in the other and it sends the wrong message to a customer, unchecked.
We use three levels of control. Every automation picks one.
In the loop: a person approves every action
The automation prepares the work, and a person says yes before anything happens.
Use this when a mistake would be costly, embarrassing or hard to undo: payments, messages to important clients, anything legal or medical. It's also the right place to start with a brand-new automation, while you learn how often it gets things right.
The trade-off: it's only as fast as the person approving. If approvals pile up, the automation becomes a queue.
On the loop: it acts on its own, and a person watches
The automation handles the routine cases by itself and sends the unusual ones to a person. Someone reviews a sample, or gets alerted when something looks off.
This is the right default for most business automations. Leave requests that clearly follow the rule get approved automatically; the ones that don't go to a manager. Invoices that match a purchase order get filed; the ones that don't get flagged.
The trick is being precise about what counts as unusual. Write it down as a rule, so everyone, including the automation, knows when to hand over.
Out of the loop: fully automatic
The automation runs end to end with no one checking each case.
Use it only when mistakes are cheap and easy to fix, and the task is well understood: sending reminders, tagging documents, moving data between systems you control. Even then, someone should own it and look at its results now and then.
How to choose
Ask two questions about the task:
- What happens if it gets one wrong? The higher the cost, the closer a person should be.
- How predictable are the cases? If most cases look alike, on the loop works well. If every case is different, keep a person in the loop.
| Mistakes are cheap | Mistakes are costly | |
|---|---|---|
| Cases are predictable | Out of the loop | On the loop |
| Cases vary a lot | On the loop | In the loop |
Control can change over time
Control isn't a one-off decision. Many automations start in the loop, earn trust, then move on the loop once the team has seen how they behave. Others move the other way after a near miss. Revisit the choice when you review the automation.
In the Builda workshop, Control is the third step of our method (Fix, Design, Control, Rank). Teams choose a level for every automation they design, and the AI coach asks them why. The answer is often the difference between an automation people trust and one they quietly stop using.