We review every AI feature we ship after ninety days. About half are still used every week. The other half had a good demo. The difference is rarely the model.
Find the question first
Kesto swimmers never asked for an AI coach. They asked why their second length was slower than their first. The stroke analysis answers exactly that, in one sentence, on the screen they already look at after a session. No chat window, no new tab.
Show your working
When the app says your catch is late on the left arm, it shows the three strokes it measured. People forgive a wrong answer they can check. They don't forgive a confident one they can't.
If a feature needs a tutorial, it is probably answering our question, not theirs.
Ship the off switch
Every AI feature we release has a named owner, a measure that tells us it is helping, and a switch that removes it without a new release. We have used that switch four times. Each time it took under a minute, and nobody outside the team noticed.
None of this is clever. It is the same product work as always, with a new kind of part.
Hannah Reid is Data & AI Lead in Sydney. Case study: Kesto Swim
