Imagine a florist adding AI to an internal tool. The home screen says only “Ask me anything.” A new colleague wants to rewrite a bouquet description but does not know which facts to provide or whether the tool will publish it. Our redesign sketch keeps chat while making the first step visible. All interfaces shown are original concepts, not product screenshots, and no user-test results are claimed.

Replace the greeting with: “Draft product descriptions from verified flower information, for your review before use.” Microsoft’s HAX guidance recommends explaining capabilities and using examples of inputs or outputs to build understanding.[1] For our fictional florist, we also propose a clear boundary: the feature does not verify stock or publish automatically. The explanation must match the actual implementation, not serve as a generic disclaimer.

Place three editable starting actions above chat: “Draft a description,” “Shorten the text,” and “Adjust the tone.” Each explains its output and allows the employee to change the goal. HAX also recommends making AI services easy to invoke when needed.[2] These are our design proposals, not evidence that three buttons suit every team. Keep free-form input available for legitimate needs outside the menu.

After Draft a description, show flower facts, length, and tone, followed by a reviewable preview. Identify missing facts instead of silently inventing them. Mark the preview as a draft requiring checks so polished formatting does not imply guaranteed accuracy. Another HAX guideline addresses realistic expectations of errors.[3] We therefore propose retaining entered facts beside generated wording for comparison, rather than hiding the input behind the conversation.

Compare the original and sketch in a small trial where actual staff complete the same authorized task. Observe whether they choose an appropriate starting point, understand needed inputs, notice facts requiring verification, and know how to revise or leave. Record where they get stuck, not just chat counts. This trial has not been conducted. Evidence of task completion should decide whether the redesign helps, not the claim that it contains AI.