There is a strong pull right now to add a model to things that do not need one. Some of it is genuine curiosity and some is that it is easier to fund.

Reach for something else when

• The rule is deterministic and known. Tax brackets, opening hours, eligibility thresholds.
• The cost of being wrong is legal, financial or safety-critical and there is no human in the loop.
• You cannot explain the decision to the person it affects, and you are required to.
• You do not have data to evaluate against, in which case you will ship on vibes and find out later.

Probabilistic tools are for problems that were already probabilistic.

Ambiguous language, messy documents, fuzzy matching, ranking — those were never solved cleanly by rules and are where these models genuinely earn their place.