The useful question is not “Where can we add AI?” It is “Which task is worth improving, and how will we know?”
Choose a bounded task.
Start with a defined input and an observable output: summarising an internal document, helping retrieve information or preparing an item for review. Broad ambitions are easier to evaluate when broken into specific workflows.
Keep a person in the right place.
Decide what the system can do automatically and which actions require review. Consider the consequences of an incorrect answer, not only the appeal of a successful demonstration.
Evaluate with real examples.
Use representative cases and difficult exceptions to judge quality, time saved and failure modes. The goal is a workflow that is useful in practice, with clear limits and ownership.