Start with one stuck workflow.
“Use AI for prospecting” is too broad to build or test. Pick a moment where the work stops moving.
Name the delay.
An interested reply sits in an inbox because someone has to find the property and reconstruct the prior conversation. A call ends with a promise that never becomes a task. A deal sheet is rebuilt because an assumption changed.
Each is a concrete starting point. You can identify the input, the missing evidence, the person making the decision, and the result they need.
Follow the work all the way through.
A draft is an intermediate result. The work may still require approval, a sent message, a confirmed meeting, or a corrected record. Design those steps together and keep a receipt of what actually happened.
This is also how you find the places where automation should stop. If an action commits someone, discloses sensitive information, or changes a financial position, the review step needs to be part of the workflow from the beginning.
Make the next attempt better.
Record the corrections. If the same missing context forces a reviewer to rewrite every draft, fix the input. If the action never happens, fix the handoff. A single working loop teaches more than a collection of disconnected demos.