Old elevators had operators. Modern elevators have buttons.
That is a useful picture of good automation. We did not keep a person manually controlling the motor “for safety.” We automated the routine work and left people a small, legible control surface: choose a destination, hold the door, stop in an emergency.
Too many AI automations do the opposite. They remove the buttons, add a progress spinner, and call the result autonomous.
The elevator handle
A handle is any point where a person can understand and redirect the system without dismantling it. It might be a preview before sending, a pause button during a batch job, an editable plan, an undo window, or a queue of uncertain cases.
The point is not constant supervision. Constant supervision turns automation into a screensaver for the employee who still carries all the responsibility.
The rubber-stamp trap
“Human in the loop” often means showing someone 500 model decisions and asking them to click Approve. By decision 80, the person is not governing an AI system. They are a biological API rate limit.
Human attention should go where human judgment changes the outcome: unfamiliar cases, irreversible actions, policy conflicts, and decisions with asymmetric harm. Routine, easily checked work can proceed automatically.
Google's People + AI guidance on graceful failure emphasizes returning control to users and giving them ways to manage errors. The idea sounds obvious. Many products still discover it immediately after their first un-cancelable batch operation.
What a handle contains
A useful intervention point carries enough state to act:
- what the system is trying to accomplish,
- what it has already changed,
- what it plans to do next,
- why it stopped here, and
- which corrections can safely resume the work.
“Something went wrong” is not a handle. Neither is dumping a raw reasoning trace. People need a compact operational summary and actions that preserve completed work.
Build for steerability
Test the interruption path as carefully as the happy path. Pause halfway through. Change the destination. Reject one item in a batch. Revoke permission. Close the laptop and resume tomorrow. Verify that the system does not repeat irreversible actions while finding its place.
The goal is not to keep a human staring at the loop. It is to keep the loop reachable by a human. Great automation disappears when things are ordinary and presents a sturdy handle when they are not.