Severitymedium
SymptomA routine runs on time, every time, and produces work that still needs the same manual correction it always did.
CausePromotion to a routine before the manual version was verified, so the schedule automates the ambiguity.
DetectionCorrection time per run does not fall as the routine matures.
FixPromote only along the ladder of one-time task, corrected task, saved skill, tested routine, with a successful run required at each step. Expand scope only after at least three consecutive runs with source-linked output, no unauthorised actions, predictable cost, recoverable failures, and less human correction time than the original workflow. A Bot that produces more activity but creates more review work has not automated the job.