Start with something boring
The best first AI projects are rarely exciting. Somebody retyping the same details every day. A search that never finds the right file. A pile of documents that need sorting.
Make a list of the jobs that eat staff time, follow a pattern and have an obvious "done". Ignore anything where nobody can describe what a good result looks like. If you cannot say what good looks like, you cannot test it.
Look at the data before the technology
AI needs something to work with. Where does the information live, how tidy is it, and who is allowed to see it? A lot of promising ideas die here, because the data is scattered across five systems, or the application where the work happens has no way of connecting to anything.
Expect the data and integration work to be bigger than the AI work. It usually is.
Run a small pilot, and measure it
Pick one task, one team and one number: minutes saved per request, say, or the share of answers that are right on a checked sample. Run it for a few weeks with a person checking the output, then compare it with how you work today.
A modest, honest result beats a polished demo every time.
Decide who looks after it
Somebody has to own the thing afterwards: watching the cost, dealing with the odd bad answer, updating it when your policies change. Settle that before you start. Projects with no owner quietly stop being used.