Most AI initiatives don't fail because the technology doesn't work. They fail because nobody answered four questions clearly enough before the budget was approved.
1. What outcome actually matters?
Not the use case, the outcome. "We're piloting an AI chatbot" is a use case. "We want to cut first-response time on support tickets by 40%" is an outcome. If the initiative can't be stated as a measurable change in the business, it isn't ready for approval yet, it's still an idea.
2. How will success be measured, and by whom?
Someone needs to own the number before the project starts, not after. If the answer to "who's tracking this" is vague, that's a signal the initiative doesn't have a real owner yet, just enthusiasm.
3. Which dependencies could actually block delivery?
Data access, integration points, a vendor's own roadmap, a compliance review, these are the things that quietly stall initiatives for months. Naming them before approval doesn't guarantee they won't happen. It means you'll recognize them the moment they start, instead of three months in.
4. What decision has to be made right now, versus later?
Not every open question needs an answer today. The discipline is separating the one or two decisions that genuinely block progress from the dozen that can wait. Trying to resolve everything up front is its own way of never starting.
Why this matters more than it looks like it should
None of these four questions are technical. That's the point. The initiatives that stall or quietly get abandoned six months in almost never fail on the model, the data pipeline, or the integration. They fail because the outcome was never pinned down, ownership was assumed rather than assigned, or a dependency that everyone half-noticed turned out to be the whole project.
Running an initiative through these four questions before approval takes twenty minutes. Recovering an initiative that skipped them takes months.