Shipping the first useful version of AI.

AI / ML · 05.05.2025 · 8 min read

Every product team is under pressure to add AI. The pressure is real but the direction is often wrong. Teams start with the technology and work backwards to a use case, rather than starting with a genuine user problem and asking whether machine learning could help.

The useful question is not "where can we add AI?" It is "where does the user face a decision that is slow, error-prone or cognitively expensive — and where do we have enough data to model that decision reliably?" The answer is usually much narrower than the initial ambition.

We have seen this pattern consistently: the AI feature that gets built in the first sprint is a demonstration of capability. The AI feature that users actually rely on is built in the third iteration, after the team has learned what the model is genuinely good at and designed the interface to expose that specific strength.

"The AI features worth shipping are the ones that make a decision the user would have made anyway — just faster."

Human-in-the-loop design is not a compromise. It is the correct starting position for almost every AI feature in consumer and enterprise products. The model surfaces a suggestion; the user confirms or corrects it; the system learns. This loop creates trust, generates training data, and gives the team a clear signal about where the model needs improvement.

The AI features worth shipping are the ones that make a decision the user would have made anyway — just faster, or with more information, or with less cognitive overhead. That is a narrower brief than "AI-powered everything", but it is a brief that leads to products people actually use.