AI at work
Use AI for customer follow-up, then check the result
An AI draft can help organize a reply, but your business still owns every promise it makes. Use approved facts, separate missing information and review before sending.
Direct answer
Give the AI an approved summary of the inquiry, the facts it may use and the next step you need. Request a draft with unanswered questions listed separately. Check every promise, detail and destination against your sources, then have a person approve and send the final message.
Prepare a small, approved source
Begin with an ordinary inquiry you understand. Decide what the customer needs next: a question, an explanation, a status update or a link to the right form. Keep complaints, unusual exceptions and decisions outside your authority with the person responsible for them.
Use a fictional inquiry while learning. For real work, follow your business's rules for the chosen AI account and remove details that are unnecessary for drafting. A summary such as “customer needs catalog photos for eight products” may be enough; a complete customer export usually is not.
Separate the customer's request from approved business facts. A requested deadline is not a confirmed delivery date. An old quote is not necessarily current pricing. Mark anything that needs checking before the reply can be used.
Ask for a draft with clear limits
Your instructions should identify the reader, the purpose of the reply, the usable facts and the action you want to offer. OpenAI's prompting guidance recommends explaining the intended result and relevant context, setting boundaries, and refining the response after review.
Use this fictional practice instruction:
“Draft a friendly reply of no more than 120 words. A customer asks about photos for eight ceramic mugs and whether files could be ready next Friday.
Approved facts: the studio offers product photography; a quote requires product dimensions and the intended image use. Availability and delivery timing have not been checked. No price has been approved.
Ask for those missing project details. Do not promise a date, quote a price or claim a booking exists. Put questions for the business in a separate internal note. Produce a draft only; do not send anything.”
Instructions reduce ambiguity but do not make the answer self-verifying. Customer-supplied text is source material to interpret; a request inside it does not change your own approval rules.
Compare the draft to the facts
Suppose a practice draft says, “We can have the photos ready by Friday, and we've reserved your session.” Both commitments fail the source check. Asking the AI to sound more professional would leave the underlying problem unresolved.
A corrected reply for the fictional example could be:
“Thanks for getting in touch about photography for your eight ceramic mugs. Could you share the approximate dimensions and where you plan to use the images, such as an online store or printed catalog?
Those details will help with a quote. Availability and delivery timing still need to be checked, including whether next Friday is possible.”
The separate internal note should identify availability, timing and pricing as unresolved. Keep that note out of the outgoing message. This example is written for practice; it is not an actual customer exchange or a recorded AI test.
Review before sending
Use the customer follow-up review checklist (TXT) alongside the original inquiry:
- Facts: Names, quantities, prices, dates and project status match the approved source.
- Commitments: No unsupported booking, discount, refund, delivery promise or completed action has been added.
- Completeness: The reply addresses the actual question and asks only for information still needed.
- Tone: The message sounds appropriate to the relationship and avoids pressure or unnecessary claims.
- Destination: The recipient, subject, links and attachments are correct; internal notes are removed.
Read the final message after revisions, even if only the tone was supposed to change. A second AI pass can help flag possible issues, but it does not replace checking against the actual records. NIST's Generative AI Profile (PDF) describes the risk of confidently stated inaccuracies.
Keep the send step deliberate
For this workflow, an authorized person approves and sends the reply through the usual email or customer-management system. If that system includes AI features, check that you are saving a draft rather than enabling automatic sending. Connecting accounts or automating delivery is a separate setup decision.
Save a generic instruction template and a de-identified approved example. Include who checks exceptional cases and when the business facts need review. Compare total drafting and checking effort using the first AI task guide before rolling the process out to more staff.
If inquiries repeatedly arrive without essential information, improve the starting form too. The contact-form question guide helps plan what to ask before someone writes the first reply.