1. Agree the boundaries first
Choose an approved tool and account, such as a business-approved ChatGPT or Microsoft Copilot workspace. Confirm your organisation's rules before use. The exercise below is entirely fictional: do not replace the details with live customer information. Work in pairs so one person drafts and the other reviews. Allow roughly 20 minutes for the first attempt and discussion.
2. Give the model a task, context and limits
Copy this fictional brief into the approved tool. A useful prompt gives the model a clear job, the available facts, the required format and instructions for missing information. It should not be treated as a security boundary or a guarantee that the model will obey.
You are helping draft a customer-service email for a fictional office-supplies company.
Facts: A customer ordered 10 desk lamps. The order is delayed. Our team is checking the delivery date. We do not yet know when it will arrive.
Task: Write a polite reply of no more than 100 words. Acknowledge the delay, explain what is known and say we will provide an update when we have confirmed information.
Do not invent a delivery date, a refund policy or compensation. After the draft, list any information a human should confirm before sending.3. Review the answer against the brief
The reviewer should check each factual statement, any implied promise and the tone. Mark unsupported details even if they sound plausible. Confirm that the output meets the length limit and does not imply an action the business has not approved. Ask whether you would be comfortable owning the message if the customer relied on it.
- Does every claim come from the brief?
- Has it invented a date, policy or commitment?
- Is uncertainty clear and useful?
- Does the message need editing before it could be sent?
4. Change one thing and compare
Ask for a second version with a different audience or tone while keeping the facts fixed. Compare the outputs against the same checklist. Notice whether changing the style also changes the meaning. Try the same prompt again: variability is a useful reason to keep the human review step. Record which instructions helped and where review was still necessary.
5. Turn the exercise into a workflow decision
Discuss whether the task is frequent enough to justify further work. Estimate drafting and review time together; a faster draft may be less useful if it takes longer to correct. Before trying real work, agree approved data, templates, reviewer responsibilities and a way to report errors. A team workshop can extend this exercise into a role-specific practice session, and a separate implementation project can address integrations.
Practical takeaways
- Use approved tools and fictional or authorised examples.
- A good prompt does not replace checking the output.
- Measure review effort as well as drafting speed.
- Agree a safe workflow before using real customer information.