Systems that do the work

Rule-based automation follows paths specified in advance and stops at anything unanticipated. An agent can interpret unstructured input and handle variation, which matters because the exceptions are usually where the human time actually goes.

Those with high volume, clear rules, and a tolerable cost of error — and where someone can describe the process precisely. A process nobody can articulate is a mapping problem first, and automating it would encode inefficiencies rather than remove them.

It depends on consequence and reversibility, not on model confidence. Actions that are hard to undo — external messages, payments, record changes — warrant approval; routine, reversible work generally does not, and requiring it everywhere trains people to approve without reading.