Do not use a brain where a light switch would work. Answer six quick questions and this tells you whether the first version should be a plain rule, an AI tool, a hybrid, or nothing yet.
1. Is the trigger clear? (a date, a count, a threshold, a missing field, a status change)
2. Is the hard part messy human input? (language, notes, images, judgment)
3. Does the output touch customers, money, safety, legal, payroll, medical, or employee records?
4. Can a human approve before it acts?
5. Is a mistake reversible?
6. Does this happen often?