When a feature should not use AI
Part of designing AI products well is knowing where a model does not belong. Four signs we look for before reaching for one.
We build AI products for a living, and one of the most useful things we do for clients is talk them out of AI in the wrong places. A model adds cost, latency and uncertainty. It earns its place only where it does something a simpler system cannot.
The answer is already known
If a rule, a lookup or a well-designed form produces the right result every time, use it. Determinism is a feature. People notice when a button that should always do the same thing does something slightly different on Tuesday.
Mistakes are expensive and invisible
Where an error is costly and hard to spot, such as a figure in an invoice or a field in a contract, a model should suggest and a person should confirm. If the design cannot make that review fast and natural, the feature is not ready for AI yet.
Speed matters more than nuance
Some interactions need to feel instant. A search box that waits two seconds for a model to rephrase the query feels broken, even if the results are better. Often the right design is a fast conventional result first, with the AI layer arriving after.
Nobody asked for it
The most common sign is the simplest. If the feature exists because AI is expected rather than because a user is stuck, it will be ignored. We start from the moment people struggle, and only then ask whether a model is the best way through it.
The best AI feature is often the one you decided not to build, so the right one gets all of your attention.
