ChatGPT for Customer Service: 12 Prompts, Team Setup, and Where It Stops

ChatGPT is a good writing partner for a customer service team. It drafts replies, calms down a tense email, summarises a long thread, and turns a policy document into FAQs and saved replies in minutes. What it does not do is answer your customers. It works for the person typing into it, and it knows what that person gives it. This guide covers the prompts that work, how to set ChatGPT up for a support team, and where you need something else. We build Evoriqa, an AI agent that answers customers from a business's own content, so the last part is about us; here is how we keep comparisons honest.

What ChatGPT does well in customer service

The jobs where ChatGPT earns its place all end with a person reading the result.

• Drafting replies. Give it the customer's message and the facts, and it writes a clear, polite answer you can edit.
• Rewriting for tone. Shorter, warmer, more formal, or plainer English for a customer whose first language is not English.
• Summarising. A long email chain or a call transcript becomes three lines and a next step.
• Writing help content. Policies, product notes and old tickets become FAQ answers, help articles and saved replies.
• Translating. A reply in Spanish or German, checked by someone who reads it before it goes out.
• Practising hard conversations. It can play an upset customer so a new team member can rehearse.

The common thread is that a person stays in the loop. ChatGPT produces a draft, and your team checks the facts and sends it. That is where it is safe, and it is where it saves the most time.

ChatGPT prompts for customer service

Copy a prompt, replace the parts in square brackets, and paste the customer's message where it says so. Each one tells ChatGPT what it may use and what to do when the facts run out. That instruction is the difference between a useful draft and a confident wrong answer.

Replies

1. Answer a question from your facts.

You are a customer service agent for [business], which [one line on what you sell]. Reply to the customer message below. Use only these facts: [paste the policy, price or order details]. If the facts do not answer the question, say what we will check and when we will reply. Keep it under 120 words, friendly and plain. Customer message: [paste]

2. Calm an upset customer.

Rewrite my draft reply to this upset customer. Acknowledge the problem in one sentence without blaming anyone, say exactly what we will do and by when, and remove anything defensive. Do not add refunds or dates that are not in my draft. Customer message: [paste]. My draft: [paste]

3. Say no politely.

A customer has asked for [request], which our policy does not allow: [paste the policy]. Write a reply that says no clearly in the first two sentences, explains the reason in plain words, and offers [the alternative we can offer]. Under 100 words.

4. Ask for the details you need.

Write a short reply asking this customer for the details we need to help: [order number, photo, account email]. Explain in one sentence why we need each one. Customer message: [paste]

Summaries and handovers

5. Summarise a thread for a colleague.

Summarise this conversation for a colleague taking it over. Give what the customer wants, what has been tried, what was promised and by when, and the next step. Five bullet points at most. Conversation: [paste, with names and contact details removed]

6. Turn a call into notes.

Here is a transcript of a customer call. Write the call notes: the reason for the call, the details given, what the customer was told, and any follow-up we owe, with its deadline. Transcript: [paste, with personal details removed]

Help content

7. FAQs from a policy.

Turn this policy into 8 FAQ entries a customer would actually ask. Use the customer's words for the questions and keep each answer under 60 words. Use only what the policy says, and mark anything unclear as [CHECK]. Policy: [paste]

8. Saved replies from recent tickets.

From these recent tickets, find the 5 questions we answer most often and write a saved reply for each. Use placeholders like {first_name} and {order_number}. Tickets: [paste, with personal details removed]

9. A help article.

Write a help article titled "[title]" for customers. Put the steps in a numbered list, one action per step, and end with a short "If this did not work" section. Use only these notes: [paste]

Language and quality

10. Translate a reply.

Translate this reply into [language] for a customer. Keep the meaning exact, keep product names and order numbers unchanged, and use a polite, neutral tone. Then list any phrase that might not translate well. Reply: [paste]

11. Check a draft before it goes out.

Check this reply before I send it. Flag anything that promises something we have not agreed, anything that could sound rude, and any question from the customer it does not answer. Customer message: [paste]. Reply: [paste]

12. Practise a hard conversation.

Play an upset customer whose [order arrived damaged / booking was cancelled]. Stay in character, push back once on my first answer, and after five messages give me feedback on what I did well and what to change.

Two of these jobs have free tools that need no prompt at all. The canned responses generator writes saved replies for common situations, and the FAQ generator builds an FAQ page from your details. If you are writing what a chatbot itself should say, chatbot script examples has twelve templates to start from.

Setting ChatGPT up for a support team

Typing the same background into every chat gets old fast, and two people asking the same question will get different answers. Three things make ChatGPT more consistent for a team.

• Give it your material once. OpenAI's pricing page lists projects and custom GPTs on the Plus and Pro plans (November 2026). Both let you save your instructions and reference files, so a chat starts with them instead of from nothing.
• Use a shared workspace. ChatGPT Business, which OpenAI sells from 2 users, is described on its page as a secure workspace with shared company context and admin controls for users, usage and billing.
• Write the house rules. Put your tone, your sign-off, your refund limits and the phrases you never use into those saved instructions, so every draft starts from them.

Keep customer data out

Paste the question, not the person. Remove names, email addresses, phone numbers, order and account numbers, and anything about health or payments before you paste a conversation. What is allowed beyond that depends on your company's own policy and on OpenAI's terms for the plan you use.

Check every fact

ChatGPT writes fluently whether or not it is right. Before a reply goes out, a person should check any price, date, policy or order status in it against your own systems. A wrong refund amount is just as polite as a right one.

Where ChatGPT stops

Everything above has a person in the middle, and that is the line ChatGPT does not cross on its own.

• It is not where your customers are. ChatGPT does not sit in the chat window on your website, answer your business phone, or reply to messages on your business's WhatsApp, Instagram or SMS. Your team uses it; your customers cannot reach it.
• It does not keep your policies current. It answers from what is in the conversation and what it already knows. A draft built on last year's returns policy reads just as confidently as one built on this year's.
• It cannot finish the job. It can write "you're booked for Tuesday", but it cannot put the appointment on your calendar, take the order, or hand the customer to a colleague.
• It is not a support inbox. Conversations stay in the chat history of whoever had them, not in a shared queue your team works through.

To put an agent in front of your customers with OpenAI's own tools, you build one. OpenAI's developer documentation offers ChatKit, an embeddable chat interface, and the Agents SDK for the agent behind it, and both need code and a server of your own (November 2026). That is a good path for a company with developers. A small support team usually wants the agent finished.

ChatGPT vs an AI agent built on your content

The two do different jobs. ChatGPT helps the people on your team; an AI agent answers the customers.

Evoriqa's agent answers from your own pages and documents, can show the sources behind an answer in your website chat if you switch that on, and says so when your content does not cover a question. Phone answering uses your own Telnyx or Infobip number, billed by that provider.

Using both

Most teams that get real value from AI use both kinds. ChatGPT helps the team write, summarise and practise. An agent on the front line answers the repeated questions at any hour, books what can be booked, and a customer who asks for a person reaches your team with the whole conversation in front of them in the inbox.

If you want to see what that looks like for your own business, paste your website into the preview and watch an agent answer from your own pages before you sign up. Plans start free, and pricing has every plan.

Frequently asked questions

Can ChatGPT replace a customer service team?

No. It makes a team faster at writing and summarising, but it does not answer customers itself, and its drafts need checking. An AI agent built on your own content can answer the routine questions directly, and your team takes the conversations that need judgement.

Is it safe to paste customer conversations into ChatGPT?

Remove personal details first: names, email addresses, phone numbers, order and account numbers, and anything about health or payments. What is allowed beyond that depends on your company's policy and on OpenAI's terms for the plan you use.

Can ChatGPT answer the chat on my website?

Not as ChatGPT. OpenAI's developer tools let a developer build an agent and embed it in a website, which takes code and a server. Evoriqa gives you a finished agent for your website chat, phone and messaging channels, answering from your own pages.

What is the best ChatGPT prompt for customer service?

One that gives it the facts and tells it what not to invent. Prompt 1 above does both: it limits the answer to the facts you paste and tells ChatGPT what to say when they do not cover the question.

Where to go next

• ChatGPT for small business: 15 uses beyond customer service, and which plan fits
• Canned responses generator: saved replies with no prompt needed
• FAQ generator: an FAQ page from your details
• Chatbot script examples: twelve templates for what a chatbot says
• For businesses: what an AI agent on your own content does
• Pricing: every Evoriqa plan