For a long time, we automated what we could automate, and the list was short. The technical constraint decided for us. That constraint has largely disappeared: almost any task that goes through text can now be delegated. So we have to decide for ourselves, and we have no criterion.
We have looked at how far to delegate a decision and what constitutes a job well done. Here is the question upstream of both.
Why time saved is a poor criterion
It is the instinctive criterion, and it fails for three reasons.
It counts what can be counted. An hour saved is measurable; what you lose in gaining that hour is not. We explored this in our article on time saved: certain efforts produce their result and transform the person doing them.
It ignores the displacement effect. We reported on a survey where nine out of ten executives see no productivity gain. A likely explanation: the time saved is captured by individuals, not by the organisation. It does not disappear, it shifts—and rarely towards what one would have chosen.
It says nothing about risk. An automated task that fails silently costs infinitely more than the time it saved. We saw this with scientific code that runs perfectly while producing a false result.
Three questions that decide better
Would the error be visible? This is the most operational criterion. A task whose result can be judged immediately—a translation you read, code you run—can be automated without much risk: you will see if it is wrong. A task whose error is invisible—a summary whose sources you will not check, an extraction of figures no one will cross-reference—requires a verification that often cancels out the gain.
Does this task teach you something? Some tedious tasks are also the ones through which you come to understand a subject. Reviewing contracts trains a lawyer; marking papers informs a teacher about their class. Automating these tasks delivers the output and removes the learning. This is what we noted regarding lesson preparation.
Is someone expecting it to come from you? Some things derive their value from the fact that a person did them. A letter of condolence, a recommendation, feedback to a colleague. The content may be identical, but the value is not, because the value lay in the cost of the gesture.
One should not conclude that effort must be preserved on principle. A large share of work is repetitive, teaches nothing, commits no one, and its errors are immediately visible. Filling in a form, reformatting a spreadsheet, extracting addresses from a list. These tasks should be automated without hesitation or nostalgia, and it is precisely here that the collapse in costs delivers its best effect. The question only arises for the intermediate zone, and that zone is narrower than one might fear.
What decides in practice—and it is not this criterion
Let us be clear-eyed: within an organisation, it is almost never a reflection on value that settles the matter. It is cost.
What is automatable and cheaper ends up being automated, regardless of what one thinks of it. This is what we noted regarding the time given back to teachers: technology does not decide whether that time goes to students or justifies larger classes. The budget decides.
This means the question posed in the title is in fact a political question disguised as a technical one. What deserves to be automated depends on who captures the gain, and that is not settled by individual reasoning.
What to take away
The criterion we propose is narrower than "does it save time" and more useful.
Automate what would show its errors, what teaches you nothing, and what no one expects to come from you. All three conditions at once. That is already enormous, and it leaves intact what deserves to remain slow.
And keep in mind the question no one ever asks before automating: what will I do with the time thus saved? If the answer is "something more urgent", then the gain does not exist. It has simply changed its name.