As long as AI merely responded, the question of delegation remained theoretical. With agents that execute tasks and now make payments, it becomes real. We already delegate a great deal to machines without thinking about it. What changes is the nature of what can be delegated.
Let's start by setting aside a false assumption: delegating a decision to a machine is not new and is not bad in itself. A thermostat decides when to heat, a cruise control decides when to accelerate, a spell checker decides what counts as an error. No one is bothered by this.
First criterion: reversibility
This is the most operational one. A decision that can be undone at no cost is a good candidate for delegation. An irreversible decision is not.
A poorly chosen route costs ten minutes. An unnecessary purchase can be returned. An email sent to the wrong person cannot be taken back. An unsuitable medical treatment even less so.
This criterion explains why automatic replenishment of everyday products bothers almost no one, while the same mechanism applied to an investment immediately raises concern. It's not the amount that matters, it's the possibility of going back.
Second criterion: verifiability of the criterion
Delegating means specifying what a good decision looks like. Yet some decisions can be expressed in explicit criteria, and others cannot.
"Buy this specific model when it drops below a hundred euros" is a complete criterion: you can check afterwards whether the agent did well. "Pick me a gift my sister will like" is not, because the real criterion is implicit, contextual, and probably unknown to you yourself until you see the object.
When a criterion cannot be formulated, delegation transfers not execution but judgment. And a judgment delegated to a system that knows neither your sister, nor your history, nor what you mean by that gift, produces a result that is statistically acceptable and personally empty.
This is the most important and most neglected criterion. Delegating without being able to evaluate what comes back is not delegating but abdicating. This is what we highlighted regarding learning to code: you cannot supervise what you do not understand. And the problem worsens over time, since delegating a skill ends up eroding it, as research on cognitive delegation suggests. Today's delegation can make tomorrow's verification impossible.
The trap of the slippery slope
No one decides one morning to entrust their important decisions to a machine. The slide happens through an accumulation of small, reasonable steps.
You first delegate search, then comparison, then pre-selection. At each step, the gain is real and the risk low. Then one day, the final decision concerns three options presented by the system, whose selection from a thousand you do not understand. You feel you are deciding, but you are choosing from a menu someone else has composed.
This mechanism is well known in interfaces, and it is nothing specific to AI. What changes is the scale of the filtering done upstream and the opacity of the criteria, as we recalled regarding the impossibility of neutrality.
What this requires
A simple practice: distinguish what you delegate by choice from what you delegate by default.
Delegating by choice means having decided that a task does not deserve your attention, knowing the criterion and being able to verify. Delegating by default means letting the tool do it because it offers to, without having thought about it.
The former is a gain in freedom: you devote your attention to what matters to you. The latter is a silent loss, because you will never know what you would have decided otherwise.
There is no universal right answer to the question in the title. There is one question to ask yourself regularly, and only one: am I delegating to free up my attention, or to avoid thinking? Technology cannot answer that, and that is rather reassuring.