An AI demo shows a polished customer reply. What you need to know is whether you could use the same tool to answer your own customers—and what you would still need to check.
Consider this example message: “My order arrived damaged. Can I get a refund?” You want help writing a considerate acknowledgement. You haven’t checked the order or decided whether a refund is due.
That small task gives you a useful way to assess a new tool. A good review should help you answer four questions before you hand over the work.
1. Can I do this in the account I have?
Check which app and plan the reviewer used. A demonstration inside a company’s connected customer-service system may work differently from the chat window you can open. Workplace administrators may also control which features are available.
For this task, start by drafting a reply in chat. You don’t need to connect your inbox or give the tool permission to send messages. If the review depends on a feature you can’t find, check the product’s availability page and your own account before following the rest of its instructions. Access through a developer service doesn’t establish access in the everyday app; region and rollout timing can also matter.
The useful decision is specific: can this account help you prepare a reply? A review that demonstrates automatic sending hasn’t yet answered that smaller question for you.
2. What happens to the information I provide?
“Not used for training” and “not retained” answer different questions. A service may keep a message for a period without using it to train its models. Privacy terms can also differ between a personal account, a workplace account and software built with a developer service.
For instance, OpenAI says GPT-6 Astra supports Zero Data Retention for eligible API customers. The API is the service developers use to connect software to a model. That qualified statement doesn’t establish a no-storage rule for every ChatGPT conversation. OpenAI’s Astra documentation
Use the short example message above for your first attempt. Before pasting an actual complaint, check your organization’s rules and the privacy documentation for the account you would use. Establish who can access or review the message, what the service keeps, and how deletion works. Removing a name may not be enough if the remaining details identify the customer.
If you can’t establish that the information belongs in that service, keep the exercise limited to example material. A vague claim that a tool is “private” doesn’t resolve the decision.
3. What happens if the task stops or changes?
A successful demonstration may leave out interruptions. Look for what happens when the tool pauses for approval, refuses part of a request or changes the model doing the work. A pause is worth understanding before you click through it: check the proposed action and whether it matches what you asked for.
For the customer reply, keep the task limited with a prompt like this:
Draft a short acknowledgement of this message: “My order arrived damaged. Can I get a refund?” Express concern about the damaged order. I haven’t checked the order or refund policy yet, so don’t promise a refund, replacement or response deadline. Prepare the wording for me to review. Don’t send anything.
If the tool stops, inspect what it has produced. A paragraph that trails off is unfinished, even if its opening sounds good. If it asks for your refund policy, you can keep this attempt to an acknowledgement; deciding eligibility is a separate task that needs the relevant facts. Don’t supply private order details just to get past an interruption.
Note any model change or limitation alongside the result when comparing tools. Save useful work somewhere you control before continuing a longer task. Safeguards and completion messages don’t establish that the answer is correct; you still need to read what the tool actually wrote.
4. What still needs my approval?
A considerate reply can contain a promise you didn’t authorize. Look closely at sentences such as “We’ll refund you within three days.” In this example, neither the refund nor the deadline has been established. Remove that commitment before considering the draft usable.
Writing, sending and issuing a refund require different decisions. A chat-only draft lets you inspect the wording before you use it. If you later connect a tool to your inbox or other business software, check its actual permission controls. A prompt saying “don’t send” states your instruction; it doesn’t prove the app technically prevents sending.
Keep your approval around actions that send or publish information, spend money, delete files, change settings, upload private material or make commitments. For the reply itself, confirm the facts and promises, then check the recipient and any attachments in the email app before sending. If a tool creates a separate file, open that file where you will use it rather than relying on its description of what it made.
Our guide to checking an AI answer offers a routine for checking claims in a draft.
You may finish this exercise with a narrower use for the tool than the demo suggested: helping you word an acknowledgement while you handle the customer record and refund decision. That’s a useful result. Keep the prompt, the checked draft and a note of anything you had to correct. You’ll have something concrete to judge the next time you consider giving the tool more responsibility.
Sources used for this article
- OpenAI: GPT-6 Astra — eligibility for API Zero Data Retention.
