automation

In depth

The three conditions

Automating a task assumes three things, and forgetting one of them explains most failures. The task must be capable of being described, its cases must be capable of being listed, and its result must be capable of being checked, failing which the machine carries out, with fine consistency, something that no one wanted. What generative models change is real: tasks that could not be written as rules become automatable, which widens the field considerably. But the widening is paid for in guarantees, because a chain of rules fails visibly whereas a model produces a wrong result with the same assurance as a right one. Checking therefore becomes the limiting factor again, and no longer the capacity to produce.

The line with augmentation

The honest dividing line with augmentation lies neither in the vocabulary nor in the technology, but in two criteria. The first is the place of judgement: where that judgement adds nothing, removing it is cheaper and more straightforward; where it constitutes the value of the work, removing it destroys what one was trying to produce. The second is the price of an undetected error: when an error is cheap and shows up at once, automation is easy to take on; when it is costly and silent, keeping a person in the loop is a matter of design, not of sentiment. Between the two lies the most frequent solution and the least announced: automate the segments and keep the decision, which gives a system where the human steps in at one chosen place rather than at all of them.

Four traps

Four traps recur with regularity. The first is to automate a bad process, which does not remove its defects but industrialises them, at a rate where no one catches them any more. The second is to steer on the volume automated rather than on the result obtained, a convenient indicator that rises even when quality falls. The third is the erosion of skills: when the machine handles the ordinary cases, people see nothing but the exceptions, that is to say the hardest cases, without practising any longer the simple cases that kept their judgement sharp. The fourth is the residual cost, which people forget to count: an automation has to be maintained, it ages along with the interfaces that surround it, and it demands work again at every change around it.