When I look at an AI feature through service design, I put the agent on the same map as the people and systems around it. What work can it do on its own? What does it need from the people upstream? Who can act on what it produces? Those questions tell me whether the feature actually changes the way work gets done.
Take a hypothetical learning service. An agent could assemble progress reports so teachers spend less time pulling information together. If the report goes on to recommend additional paid lessons, the stakes change. I would give teachers a chance to see the evidence and decide whether the recommendation is justified. The approval has to sit in the delivery flow, where a report can be held for review before it goes out.
I would plan the work in three steps. First, agree on the problem, the desired outcome, and the limits. Then decide which parts of the report the agent can prepare and which decisions require a person. Finally, test the riskiest hand-off with teachers before using the workflow more widely. Written guidance can shape the tone of a draft. A review gate decides whether the draft gets sent. Both belong in the design.