A good first project starts with one task that already happens often enough to be painful. The goal is not to automate a whole department in one pass. The goal is to prove that a small, repeatable loop can be trusted.
That means clear inputs, expected outputs, evidence trails, and a way for a person to review or stop it. The early win is usually not glamour. It is one recurring job that stops taking attention every week.
If the build survives real data, real delays, and real edge cases, it earns the right to become permanent infrastructure. If it does not, the business still leaves with a precise map of the process and the weak points.