Repeatable, defined hard

Everybody agrees a task should be repeatable before automating it, and almost nobody tests whether it is. Two instruments, and they take an afternoon between them.

The four questions. Can you write the rule in a way that two people would apply identically? Does the rule cover every case seen in the last month, including the awkward ones? When the rule does not fit, is there a defined path, or does somebody improvise? And has the rule changed in the last year, in which case what changed it, and will that thing happen again?

Then the ten-example audit, which is where the answers get tested. Take ten real, recent instances, chosen at random rather than chosen as representative. Apply the written rule to each yourself. Count how many the rule handles completely.

Under eight out of ten and the task is not ready. That is not a failure, it is a finding: you have discovered that a process everybody believed was standard is actually being held together by a person making calls.

Choosing the examples at random rather than picking good ones is the part people skip, and it is the part that makes it an audit rather than a demonstration.