ChatGPT prompts for customer service replies that do not sound like a robot

The short version: ChatGPT can cut the time it takes to write a customer service reply from ten minutes to under two, but only when you paste in the actual complaint and your actual policy first. A copy paste generic apology gets noticed and makes things worse. The prompts below are built to produce a specific reply you edit for tone, then send yourself, never one you send straight from the model.

Why this is a good use of ChatGPT, with one condition

Most customer service replies follow a shape: acknowledge, explain, resolve, close. ChatGPT is fast at that shape. What it cannot do reliably is know your refund policy, your delivery times, or what happened to a specific order, so every prompt below asks you to paste those facts in rather than trusting the model to know them. The condition is simple: never let it invent a policy, a discount, or a timeframe you have not given it, and always read the draft back before it goes anywhere near a customer.

Prompts for refund and return requests

• Prompt: "A customer is asking for a refund on [product] because [reason they gave]. Our policy is [paste your actual policy]. Write a reply that confirms whether they qualify, in plain language, no corporate phrasing."
  What to change: the product, their stated reason, and your real policy text.
• Prompt: "This customer is outside our 30 day return window by [number] days. Write a reply that explains the policy without sounding like a form letter, and offer [alternative, such as store credit] if that is something we can do."
  What to change: the number of days over and whatever alternative you are willing to offer.
• Prompt: "Write a short reply confirming a refund of [amount] has been processed and will appear in [timeframe], based on our payment provider [name]."
  What to change: the amount, timeframe, and provider name, since these details differ by payment method.
• Prompt: "A customer wants an exchange, not a refund, for [product] in [size or variant]. Write a reply confirming next steps and what they need to send back."
  What to change: the product and variant, and your actual return shipping process.

Prompts for complaints and unhappy customers

• Prompt: "Here is the customer's message: [paste full message]. Write a reply that acknowledges their specific frustration in the first line, not a generic 'sorry for the inconvenience', then explains what we are doing about it: [paste your actual resolution]."
  What to change: the full message and your real resolution. Copying the customer's own words back to them, briefly, shows you read it.
• Prompt: "This customer has emailed us three times about the same issue. Here is the history: [paste thread]. Write a reply that does not repeat what we already said, moves the situation forward, and offers one concrete next step."
  What to change: the thread history. Repeating the same non answer is the single fastest way to lose a customer for good.
• Prompt: "Rewrite this draft reply to sound less defensive and more like we are on the customer's side, without admitting fault we do not owe: [paste draft]."
  What to change: your own draft. This is useful when a reply you wrote while frustrated reads as combative.
• Prompt: "A customer is threatening to leave a bad review over [issue]. Write a reply that addresses the issue directly and offers [your actual resolution], without mentioning the review threat at all."
  What to change: the issue and your resolution. Never reference the review threat back to them, it reads as pressure rather than service.

Prompts for shipping delays and stock issues

• Prompt: "An order is delayed by [number] days because of [real reason, such as courier backlog or supplier delay]. Write a proactive email to send before the customer asks, explaining the delay and the new expected date of [date]."
  What to change: the delay length, the real reason, and the new date. A proactive message before a complaint arrives cuts inbound queries substantially.
• Prompt: "A customer's item is now out of stock after they ordered it. Options are [list actual options, such as refund, backorder, or substitute]. Write a reply presenting these clearly and asking which they prefer."
  What to change: the real options available for that specific product.
• Prompt: "Write a short apology for a delay with no specific compensation offered, because our policy on this issue is [paste policy, or 'we do not offer compensation for standard delays']."
  What to change: the policy. Do not let the model invent a discount code or refund you have not authorised.

Prompts for reviews and public feedback

• Prompt: "Here is a one star review: [paste review]. Write a public reply that is calm, takes the specific complaint seriously, and invites them to contact us directly at [contact method] to resolve it."
  What to change: the review text and your real contact method. Public replies get read by future customers, so keep it short and non defensive.
• Prompt: "Here is a five star review that mentions [specific detail]. Write a short public thank you reply that references that detail, not a generic 'thanks for your kind words'."
  What to change: the specific detail mentioned. Referencing it shows a real person read the review.
• Prompt: "A review makes a factual claim about our product that is not correct: [paste claim]. The correct fact is [paste correct fact]. Write a polite public correction that does not accuse the reviewer of lying."
  What to change: the claim and the correct fact. Tone matters more than accuracy here, since a defensive correction reads worse than the original review.

Prompts for internal tone and templates

• Prompt: "Take this reply and rewrite it at three lengths: a two sentence version for a quick chat reply, a paragraph version for email, and a longer version with full context for an escalated case: [paste base reply]."
  What to change: the base reply. Different channels need different lengths, and this saves rewriting from scratch each time.
• Prompt: "Turn this one off reply into a reusable template with brackets for the parts that change each time: [paste reply]."
  What to change: the specific reply you want to templatise. Use this once you have a strong reply you expect to send in similar form again.

A weak prompt versus a strong prompt

Weak: "Write a reply to an angry customer about a late delivery."

This produces a generic apology with a made up compensation offer, a made up delivery estimate, and language so vague it could apply to any business. If you send it as is, the customer notices instantly that nobody looked at their order.

Strong: "Order 4471 was due on the 3rd and has not arrived. The courier shows a delay due to a depot backlog, new estimate is the 9th. Our policy for delays over 5 days is a 10 percent refund of the order value, which is 6.20 pounds on this order. Write a reply to the customer, [first name], that names the new date and the refund amount exactly."

The strong version gives a real order number, a real cause, a real date, and a real figure. The reply the model produces will be specific enough that the customer believes someone looked at their case, because someone did.

Mistakes to avoid

• Letting the model invent a compensation amount. Always state your real policy figure in the prompt. A model left to guess will often offer more than you intended, or promise something you cannot deliver.
• Sending the first draft unedited. Read every reply once for tone before it goes out. AI drafts sometimes land as slightly too formal or slightly too apologetic for your brand voice.
• Copying the same reply to multiple customers with the same complaint. If ten customers are affected by the same issue, write one strong prompt with the shared facts, then personalise the opening line for each so it does not read as a form letter.
• Ignoring channel differences. A reply that works well in email can read as far too long in a live chat window. Ask for a shorter version rather than pasting the email reply straight into chat.
• Forgetting the customer's own words. A reply that never references what the customer said reads as automated even when a human wrote every word.

I have written more around this on the site: Does ChatGPT Give The Same Answer Every Time? No, And Here’s Why That Matters, What Does ChatGPT Prompt Meaning Really Involve?, How to Write Master Level ChatGPT Prompts for Business Tasks.

For the rest of the series, see the ChatGPT Prompts and Limits: 30 Guides to Fun Prompts, Photo Edits and Daily Caps.

The free ChatGPT prompt generator does the working out for you.

Related: no code (this site).

Frequently asked questions

Will customers know a reply was written with ChatGPT?

Only if it stays generic. A reply that names the order number, the specific issue, and a real resolution reads as personal regardless of how it was drafted. The tell is vagueness, not the tool itself.

Is it safe to paste customer data into ChatGPT?

Avoid pasting full names, email addresses, card details, or anything else identifying wherever you can use a placeholder instead, such as "the customer" or an order number alone. Check your own data protection obligations before pasting anything personal into any external tool.

How do I keep replies consistent with our brand voice?

Paste two or three of your best past replies into the conversation before asking for a new one, and ask the model to match that tone exactly. This works far better than describing your tone in the abstract.

Should I use ChatGPT for angry or escalated complaints?

It can help you draft a calmer first version, but a seriously escalated case, especially one involving a refund dispute or a threat of legal action, needs a human decision on the resolution before any reply goes out.

Can ChatGPT handle replies in a language other than English?

It can draft a reasonable version in most major languages, but always have a fluent speaker check it before sending, since small tone mistakes in a second language can come across as ruder or more formal than intended.