Rules first, models second: how we scope automation

Most repetitive work is rule-shaped. Use a model only where documents or judgement vary.

When a client asks for “AI automation” of a weekly report, the first question is whether the rules are known. Usually they are: copy these columns, sum those, flag anything above a threshold. That is a scheduled job, not a model.

Models earn their place where the input varies — reading a supplier’s PDF in a format nobody controls — or where judgement is needed and a human will review it.

Scoping this way keeps automations testable and cheap. Every automation still gets an exception queue and a named owner, because the interesting cases are the ones the rules did not foresee.