“Write a reply to this customer.” AI will write one. It does not know what you promised yesterday, whether you can offer a discount or why the customer is upset. Gaps in the brief can become mistakes in the reply.
Imagine handing the task to a colleague who missed the last meeting. What would they need to know to carry on?
Give it something specific to work with
Suppose a customer is asking about a delayed order. Your brief could look like this:
Draft a short reply to the attached email. The order is late and we do not have a confirmed delivery date. Apologise for the delay and say we will send an update by Thursday. Do not promise a delivery date or compensation. If you need more information, put a question for me below the draft.
That is enough for a first attempt. It also tells you what to check: the reply can promise an update on Thursday, but it cannot promise delivery on Thursday. Smooth writing can make that distinction easy to miss.
Choose the source material to fit the task. For a complaint, the relevant correspondence and current complaints procedure may be enough. Adding the entire company archive would make checking harder. Only use information you are allowed to put into the tool.
Make the next attempt more specific
If the draft is too long, say what it can leave out. If it has added a promise you cannot keep, point to the sentence and explain the limit. “Try again, but better” leaves AI guessing.
Save a brief that works, along with a good result. Next time, you can replace the source material and check whether the rules still apply. You do not have to start with an empty chat each time.
Read the reply before sending it. A clear brief can still produce a bad answer. In this example, I would pay particular attention to dates and anything we are promising the customer.