Where a rule is the right answer
- Assignment: new lead in this pipeline goes to this queue.
- Timers: nothing has happened for four days, so raise a reminder.
- Status hygiene: mark the record when the signed document arrives.
- Notification: tell the owner when a job crosses a threshold.
- Anything a new hire could execute correctly from a one-line instruction.
Where a rule stops being enough
Rules cannot read a message and understand what the customer wants. They cannot write a follow-up that references the specific hesitation the homeowner raised. They cannot decide that three separate signals together mean a job is about to go sideways.
That is the boundary. The moment the correct action depends on interpreting unstructured language or weighing several soft signals, a rule either becomes an unmaintainable tree of conditions or it becomes wrong.
How they fail differently
A broken rule fails loudly and identically. It fires on the wrong records every time, or it never fires. You find it once and fix it once.
A misfiring AI step fails quietly and variably. It produces something reasonable-looking that is subtly wrong on some inputs and right on others. That is why AI execution needs three controls a rule does not: a bounded set of actions it is allowed to take, an approval step for anything consequential, and a stored record of both what it decided and what data it based that on.
The safe division of labour
- Rules detect, route, remind and record.
- AI drafts, summarises, classifies and proposes.
- Humans approve anything that spends money, contacts a customer for the first time, or cannot be undone.
- The system stores what each of the three did, separately attributed.