Choose one useful outcome.
My starting point would be a single, bounded job: turn an approved source into an answer, prepare a status summary for review, or help someone find the right next step.
“Add AI to the business” is not a job description. “Draft a weekly account summary from these three approved systems, with links to the evidence” is something you can inspect and improve.
Write the boundary before the prompt.
Define the inputs, the expected output, the allowed tools, and the decisions that still belong to a person. For an initial version, keep the action reversible or make it a draft.
A useful first agent should be easy to describe—and easy to stop.
Use a small set of real examples.
Collect representative inputs, including a few that are incomplete or contradictory. Describe what a good response would look like, what would be misleading, and when the system should decline to act.
Show the human what happened.
Return the answer with source links, explain uncertainty, and make the next step explicit. Do not make the user reverse-engineer what the system used or whether it actually completed an action.
Then decide what to build next.
If the narrow workflow is useful, you have a clearer basis for expanding it. If it is not, the scope is small enough to change without defending an entire “AI transformation.”