The cheapest time to kill a bad AI project is before it starts. Most of the failures we see were not build problems. They were decided at the kickoff, when nobody asked the questions that would have surfaced the risk. Here are the ones that matter.
What does success look like, as a number?
If you cannot state the outcome in a figure, hours saved, conversion lifted, cost removed, you will not be able to tell whether it worked. Vague goals produce impressive demos that no one uses.
Is the data ready?
AI is only as good as the data it can see. Before committing, find out where the data lives, who owns it, and how clean it is. Data readiness, not model choice, is usually the real constraint.
What does it cost per use?
API calls, compute, and storage add up quietly. Model the cost per use and check that your pricing or savings cover it with margin. A feature that loses money on every use should be caught here, not after launch.
Who owns it, and what happens when it is wrong?
A system nobody owns rots. Decide who runs and improves it. And ask what the cost of a wrong answer is, what the blast radius looks like, and whether a human needs to stay in the loop where the stakes are high.
The cheapest time to kill a bad AI project is before it starts.
These are the first questions we ask in any consulting engagement. Honest answers up front save the budget that vague ones quietly spend.