Start with repeated friction
Ask process owners where the same work happens every week. Look for copying between systems, repeated questions, chasing approvals, or classifying incoming information. Record the steps and exceptions before discussing tools.
Separate useful from consequential
Consider the cost of an incorrect action. Drafting a summary for review is different from sending a payment or changing a customer record. The first pilot should have a manageable failure impact and a clear human fallback.
Check the connections
Confirm that the required systems expose the right access and that the underlying data is usable. A promising idea can become an integration project if the source information is inconsistent or inaccessible.
Define success before building
Measure the current workflow: time per task, number of handoffs, error rate, and exceptions. Agree what improvement would justify continuing. Compare the pilot against that baseline rather than a generic promise about AI.
Give someone ownership
A pilot needs a business owner as well as a technical implementer. Decide who reviews outputs, resolves exceptions, approves changes, and decides whether the system is ready for a broader rollout.