Use ChatGPT to prepare responses, summarize a long exchange, or identify which policy question needs a supervisor. Begin with a small authorized excerpt and keep the current ticket open beside the draft. You do not need an integration to practice this review method with fictional material.
Supply the rule that controls the answer
This original fictional case concerns a damaged desk organizer:
| Record | Supplied information |
|---|---|
| Customer message | The organizer arrived cracked. Can you replace it by Friday? |
| Policy P4 | For damage reported within 14 days of delivery, request the order reference and a clear damage photo before reviewing replacement eligibility. |
| Shipping note S2 | Standard shipping is estimated at 3 to 5 business days after dispatch. Dispatch date is not supplied. |
| Ticket facts | Delivery was four days ago. The order reference and photo are missing. |
The timing supports asking for the two required items. It does not establish eligibility, dispatch, stock availability, or Friday arrival. Keep customer statements distinct from information checked in the order system.
OpenAI's prompting guidance recommends supplying the goal and relevant context, then reviewing and refining the response. For service work, the controlling context is the actual policy version and the known case facts. Use only information permitted in your account and workspace; remove details the drafting task does not need.
Separate the customer reply from the internal question
A bounded customer draft for this fictional case is:
I'm sorry the organizer arrived cracked. Please send your order reference and a clear photo of the damage so we can review replacement eligibility. I can't confirm Friday delivery: the shipping estimate is 3 to 5 business days after dispatch, and a dispatch date has not been confirmed.
The draft states what the customer can do now. It avoids “your replacement is on its way,” which would claim an action that has not happened. It also avoids adding a refund option that P4 never supplies.
An internal note can be more explicit: “Friday delivery requested; eligibility review awaits order reference and photo. Confirm whether an authorized expedited option exists before offering one.” That question belongs to the authorized service owner, rather than appearing as a speculative offer in the customer reply.
If the customer already supplied a photo earlier in the thread, asking for it again creates needless work. Read the complete relevant exchange before requesting missing information. A summary should retain the message or attachment reference so the agent can check it.
Copy a policy-bound reply prompt
Draft a customer-service response for my review.
Customer's latest request and relevant earlier exchange: [PASTE]
Current approved policy, version/date, and section IDs: [PASTE]
Verified case facts, with source references: [PASTE]
Customer statements not yet verified: [PASTE]
Approved tone and response length: [ENTER]
Answer the customer's actual question. Use only the supplied policy
and facts. Do not invent eligibility, a refund, stock, dispatch,
delivery dates, completed actions, or an exception to policy.
Ask only for information still needed after reading the exchange.
Return a customer draft, a short claim-to-source check, and a separate
internal question for anything the policy does not resolve.
Keep internal notes out of the customer draft. Do not send anything.
For the sample, check four phrases before using the result: damage, required evidence, eligibility, and arrival date. Each needs the right status. “Reported damage” can be acknowledged; “approved replacement” cannot be inferred from it.
Resolve conflicting policy before polishing the wording
If another document says replacements are automatic, ask which policy applies to this product, purchase date, and channel. Do not instruct ChatGPT to combine incompatible rules into a friendly compromise. Record the unresolved section IDs and obtain the responsible person's decision.
The customer-service procedure guide covers clarification, resource selection, and follow-up across a ticket. This drafting task is narrower: check that each proposed promise is authorized before the agent sends the reply through the approved service system. After sending, record what was actually sent and continue the case according to the team's process.