Is there a decision or workflow to improve?
A use case should name the user, the task and the measure of improvement. If it cannot, it is a technology demonstration rather than a use case.
Is the data ready?
Check availability, quality, permission to use and stability over time. Many ideas fail here, and a short data review is far cheaper than a failed build.
Also check whether labels or ground truth exist. Without a way to judge correctness, there is no way to evaluate a model.
What happens when the system is wrong?
Estimate the cost and reversibility of errors. A wrong movie recommendation is trivial; a wrong credit decision is not. The consequence of error should determine how much human review and control the design needs.
Can it be operated?
Consider who will monitor the system, retrain it, answer user questions and pay for compute. Solutions without an operating owner decay quickly.