1. What does this workflow cost today?
In hours or headcount, stated by someone who does it. If nobody can answer, you are shopping for a problem rather than solving one — start at the expensive workflow.
2. Can a human check the output in seconds?
If not, you cannot evaluate it, catch drift, or trust it. That single property decides whether this is a safe first project or something that should wait.
3. What does a mistake cost, and who notices?
Recoverable, invisible, or reputational. Anything irreversible or customer-facing multiplies the governance work, and it is where escalation on stakes becomes mandatory rather than nice.
If nobody can say what the current process costs or how a mistake would be caught, the project isn’t ready to start.
4. Does the knowledge exist in writing?
Usually partly. The gap is the real scope, and discovering it late is why estimates are wrong — the knowledge base often is the project.
5. What must it never be able to see?
Ask in week one. Retrofitting permissions into retrieval after indexing everything is the most expensive correction in the category.
6. What would make us stop?
A result that means abandon, agreed before building. Without it the project cannot conclude, only continue — the anatomy of a permanent pilot.
7. Who owns it in eighteen months?
A name, for the evaluation, the knowledge, and the reviews. If there isn’t one now there won’t be one later, and unowned systems decay.
8. What does it leave behind?
Retrieval, logging, corrections, an evaluation harness. Prefer the project whose by-products make the next one cheap — that is how a first feature becomes a capability.