Adding a language model to a product is easy. Adding one that survives contact with real users, real budgets and real edge cases is a different job, and most of that job is deciding where the model should not be.
The features that stayed in production for me all shared a trait: a person was already doing the task, the task was repetitive, and being roughly right was good enough because a human still saw the result before it mattered.
Follow-up is the clearest win
When a lead fills in a form, someone has to read it, judge how serious it is and write a first reply. That is judgement work, repeated hundreds of times, with a forgiving margin for error.
Automating that step meant leads got an answer in minutes instead of the next morning. The model did not need to be brilliant; it needed to be immediate and never off-brand.
Where I stopped
Anything with a number attached to it — pricing, availability, medical or legal specifics — goes through normal code and a real data source. A model that guesses a price confidently costs more than it saves.
The rule I work to now: let the model handle language, let the system handle facts.