The problem
1. What exactly is the task, and who does it today? If you cannot name the person and the steps, it is too vague to build.
2. How many times does it happen, and how long does each take? Without a number, you cannot tell whether AI is worth it.
3. What does a correct result look like, and who decides? You will need that person to check the first results.
The data and systems
4. Where does the information live? A few systems you can connect to is fine. Paper and memory is a problem.
5. Who is allowed to see it? Permissions need designing in from the start, and they are harder to add later.
6. Can the systems involved be connected to, through an API, a database or a file? If not, plan for that work.
The people and the risks
7. What happens when it gets something wrong? Decide which mistakes are tolerable, which need a person to check first, and which must never happen.
8. Who owns it afterwards? Somebody has to watch it, handle bad answers and update it when things change.
9. What are staff worried about? If the people who do the work are not on board, the project will quietly fail.
Success
10. How will you know it worked? Pick one measure, such as time saved or errors avoided, before you start, and compare against it afterwards.
If you can answer most of these, you are ready for a pilot. If you cannot, a short discovery piece will answer them cheaply.