Hand the Routine to the Machine

A computer is most useful not when a person does their usual work faster with its help, but when they stop doing the repeatable part by hand.

Calculations, sorting, renaming files, checking formats, processing tables, finding duplicates and moving data can often be automated. A one-off task does not always justify building a tool. But a regular manual action gradually turns into an expensive ritual.

Automation used to mean mainly an exact sequence of commands. Such a machine does not understand intent: it carries out the instruction, including the part where the person accidentally ordered it to delete the data they needed.

Generative AI works differently. It not only follows the command but reconstructs the likely intent from the wording and the context. That makes it more flexible and at the same time more dangerous: the system can convincingly perform the wrong task, and the mistake will look like a well-written result.

This is where the formula helps: AI is not a solver but an amplifier. It increases the speed and scale of the work, but along with a good method it amplifies a bad problem statement, a hidden assumption and a careless check.

Ordinary automation needs a test set, control over units and formats, a backup copy, a comparison against a manual calculation, handling of exceptions and the ability to undo an action.

Delegating to a generative system adds context, acceptance criteria, permitted sources, examples of a correct result, a ban on inventing missing data, and a point where a human has to check and approve the outcome.

Basic programming remains useful, even though the machine increasingly writes the syntax itself. The value shifts elsewhere: break the task down, define the data and the interfaces, read the code you got back, come up with tests, and notice when the program is solving the neighbouring problem.

Studies of professional work show a jagged frontier of AI capability: inside it people do tasks faster and better, and outside it the system’s suggestion can make the result worse, because it looks more convincing than one’s own mistake. And the frontier keeps moving.

A bad process accelerated by a computer becomes a fast bad process. Sometimes fast enough that the error spreads before anyone finishes their coffee. What is worth automating is not only the action but the control over it.

Manual work is dangerous not because it is slow but because its errors are invisible. Someone who assembles a report every month by copying from four files does not get it wrong every time — roughly one time in ten. Nobody looks for the error, because the nine previous reports were correct and the procedure is considered reliable. A script in the same place gets it wrong either every time or never, and the very first run shows it. A reproducible error costs less than a random one.

So it is more sensible to start automating with the most boring thing rather than the most complex. Boring work is done without looking, and that is exactly where nobody rechecks what a person did for years on end.