Start with the work.

The most useful question is not always which model to use. Begin with a concrete task: what takes time, repeats or makes someone look for the same information again? Observe the real workflow and note the exceptions. They often reveal the constraints that matter most.

Make success observable.

Choose a focused first use case. Describe the outcome you want and gather representative examples, including difficult ones. Time saved matters, as do answer quality, ease of correction and understanding how the system works. This gives you a solid basis for assessing a prototype beyond its demonstration.