Not every product needs AI

There is real pressure right now to put AI in everything: investors ask about it, competitors announce it, and roadmaps quietly bend around it. The result is often a feature nobody asked for, stapled onto a product that had a different problem.

Start from the problem, not the model

If a rule, a lookup table, or a well-designed form solves it, that solution is faster, cheaper, testable, and explainable. AI earns its place when the input is genuinely messy: free text, images, judgement calls that do not reduce to rules, work that needs to happen at a scale no person can sustain.

The honest checklist

Before we build an AI feature we ask a few blunt questions:

• Is the task actually fuzzy, or just not automated yet?
• What does a wrong answer cost, and who catches it?
• Is there real data to ground it in?
• Can a person review the output before it matters?

If those answers are wrong, we say so, and we build the simpler thing instead.

When it is the right answer

When it fits, AI is genuinely transformative: drafting work that people only need to review, triaging piles of unstructured information, spotting patterns across more data than anyone can read. Often the way to find out is to build the boring version first. If the boring version already solves it, you have your answer, and it cost you very little.

The cost nobody budgets

An AI feature is not done when it ships. It needs evaluation sets that grow with real usage, monitoring for quality drift, spend caps so a busy day is not an expensive day, and re-testing every time the underlying model changes under you. That ongoing bill belongs in the decision from the start, because it is exactly the part the demo never shows.

Saying no to AI where it does not belong is what makes our yes worth something. The point of a product is not to contain AI. The point is to solve the problem.