AI features that earn their place in a small product

Almost every AI request I get starts the same way: someone has seen a demo and wants one. The useful question is not which model to call — it is which manual step currently costs them an hour a week.

Once the question is framed that way, the design gets much smaller. A model that drafts the first version of something a human already writes every day is worth building. A model that invents content nobody asked for is a liability, because someone now has to check it.

The part that decides whether it ships is the failure mode. A drafting feature can be wrong and still be useful, because the human is already in the loop reviewing it. A feature that acts without review has to be right, and "usually right" is not a specification you can hand to a small business.

Practically, that means most of the engineering is not prompt work. It is the boring scaffolding: keeping the request cheap, caching what does not change, making the output editable, and leaving an obvious path back to doing it by hand when the model is unavailable.

The tools that stayed in use after a month were the ones where the person could see exactly what the model did and change it. The ones that got switched off were the ones that felt like a black box, regardless of how good the output was.