The most tempting AI ideas are often the ones nobody questions. They sound plausible, everyone nods, the budget flows. Only months later comes the uncomfortable question: did it actually achieve anything? And that's when the number you'd read it from is missing.
No metric, no decision
A success metric isn't bureaucracy; it's the basis of every decision that comes later. Scale or stop? Extend or rebuild? Without a number agreed in advance, that decision becomes a matter of opinion — and whoever is loudest wins.
The metric has to come before the build, not after. Look for it in hindsight and you'll certainly find one that looks good. That isn't measurement anymore; it's self-affirmation.
An AI initiative without a success metric isn't an experiment — it's a bet, except no one knows whether it was won.
The right metric is a business number
Model quality like accuracy or F1 is a means, not an end. What counts is the effect in the process: time saved per case, cost per case, error rate in the workflow, throughput. A model can be technically brilliant and still move nothing.
A good metric is a single, clearly measurable business number that the business and management agree on beforehand. It forces honesty about what the initiative is really meant to improve.
An honest no is valuable
The real benefit of a metric defined upfront shows when an idea clearly won't reach it. Then the initiative can be stopped early and cheaply — before budget burns in a pilot with no future.
We say that no deliberately early. An expensive graveyard of pilots isn't created by bad technology, but by the missing courage to stop hopeless ideas in time.
Before the first line of code, a number should be fixed: how will we know in three months that it worked? If that number doesn't exist, the use case isn't ready yet — no matter how good it sounds.