ChatGPT for Customer Service: Prompts and Limits

TL;DR

• ChatGPT is a good assistant for the person answering tickets: drafting replies, summarizing long threads, fixing tone, translating and writing macros.
• It is a poor stand-in for your support team when customers talk to it directly. It doesn’t know your pricing, policies or account data, and it will fill the gaps with confident guesses.
• Keep personal data out of prompts, and check your plan’s data settings before pasting anything from a ticket.
• If you want customers to get answers without a person in the loop, use a bot that answers only from your own knowledge base, hands off to a human, and asks before it changes anything.

Below are the prompts we’d actually use, the failure modes to watch for, and a short guide for picking the right setup.

What ChatGPT is good at in customer service

Most support work is writing. Someone reads a message, figures out what the customer needs, finds the answer and writes it in a way that doesn’t sound annoyed. Large language models are strong at the first and last parts of that loop. They’re weak at the middle part, finding the correct answer, unless you give it to them.

That split tells you where ChatGPT for customer service pays off:

• Drafting replies. You know the answer; ChatGPT turns your three bullet points into a clear, friendly message.
• Summarizing tickets. A 40-message thread becomes five lines: the problem, what was tried, what’s still open.
• Writing and cleaning up macros. Saved replies get stale and inconsistent. ChatGPT can rewrite a batch of them in one voice.
• Translation. You can reply to a customer in Portuguese without knowing Portuguese, and read their message in your own language first.
• Tone. Softening a blunt reply, or making an apology sound less like a legal notice.
• Turning answers into help articles. If you’ve explained the same thing ten times, paste your best reply and ask for a help-center draft.

In all of these, a person stays in charge. The agent supplies the facts, ChatGPT supplies the wording, and the agent reads the result before it goes out. That’s the safest and most useful way to use ChatGPT for customer support today.

How to use ChatGPT for customer service: 8 prompt templates

Copy these, replace the parts in square brackets, and adjust them to your product. Each one gives the model the facts it needs and tells it what not to do. That second part matters more than people expect.

A quick rule before you paste anything: remove names, emails, order numbers, addresses and payment details from the ticket text. Replace them with placeholders like [CUSTOMER] or [ORDER_ID]. You can put them back when you send the reply.

1. Draft a reply from your own facts

You are a support agent for [PRODUCT], a [one-line description].
Write a reply to the customer message below.

Use ONLY these facts:
- [fact 1]
- [fact 2]
- [fact 3]

If the facts don't answer the question, say we'll check and get back to them.
Don't mention prices, dates, refunds or features that aren't in the facts.
Tone: friendly, plain, no exclamation marks. Max 120 words.

Customer message:
"""
[paste message, with personal data removed]
"""

2. Summarize a long ticket for a handoff

Summarize this support thread for a colleague who hasn't seen it.
Format:
- Problem (one sentence)
- What the customer has already tried
- What we have already said or promised
- Open questions
- Suggested next step

Don't add anything that isn't in the thread.

Thread:
"""
[paste thread]
"""

3. Rewrite a reply in a calmer tone

Rewrite this reply so it sounds calm and helpful, not defensive.
Keep every fact and every step exactly as written.
Don't add apologies for things we didn't do wrong.
Keep it under the original length.

Reply:
"""
[paste your draft]
"""

4. Translate both ways

Step 1: Translate the customer's message into English and list anything
ambiguous in the original.
Step 2: Translate my reply into [language]. Keep product names, button
labels and code in English. Use a neutral, polite register.

Customer message:
"""
[paste]
"""

My reply:
"""
[paste]
"""

5. Turn repeated answers into a macro set

Below are replies our team has sent for the same type of question.
Write one reusable macro that covers all of them.
Mark variable parts as {{placeholders}}.
Add a one-line note on when NOT to use this macro.

Replies:
"""
[paste 3–5 real replies]
"""

6. Classify and prioritize an inbox export

Classify each message below into one of these categories:
billing, bug, how-to, feature request, account access, other.
Give each a priority from 1 (low) to 4 (urgent). Urgent means data loss,
payment failure or the customer can't log in.
Return a table: id, category, priority, one-line reason.
Don't guess: if a message is unclear, mark it "other".

Messages:
"""
[paste id + text, personal data removed]
"""

7. Draft a help article from a solved ticket

Turn this solved ticket into a help-center article.
Title as a question customers would search for.
Sections: short answer, steps, common mistakes.
Use only the steps that worked in the ticket. Don't invent settings or menus.

Ticket:
"""
[paste]
"""

8. Write a “we fixed it” follow-up

Write a short message to customers who reported this bug, telling them it's fixed.
Bug: [one line]
What changed: [one line]
What they should do now: [one line, or "nothing"]
Thank them for reporting it. No marketing, no discount offers. Max 80 words.

These prompts share a pattern: give the model the facts, limit what it’s allowed to claim, and define the output format. Without the first two, you get a nicely written reply that may contain things your company never said.

Where raw ChatGPT breaks in customer support

Things go wrong when ChatGPT stops being a writing tool for your agents and starts talking to customers on its own, or when agents trust its facts instead of checking them.

It makes up your pricing and policies

A general model has never seen your refund policy, your current plans or last week’s change to shipping times. Ask it “Can I get a refund after 30 days?” and it will produce a plausible answer based on what companies usually do. Plausible and correct are different things, and in support the wrong answer creates a second ticket and sometimes a chargeback.

Giving it your facts in the prompt helps, as in template 1. But that only works when someone chooses the right facts every time.

It can’t see the customer’s account

“Where is my order?”, “Why was I charged twice?” and “Which plan am I on?” are some of the most common support questions. ChatGPT can’t answer any of them because it has no connection to your orders, billing or user database. Whatever it says about a specific account is a guess.

Privacy and personal data

Tickets are full of personal data. Pasting them into a chat tool means that data now lives in another service. OpenAI’s own help pages explain that for individual ChatGPT accounts, conversations can be used to improve its models unless you turn off Improve the model for everyone in Data controls, and that Temporary Chats aren’t used for training. For its business offerings (ChatGPT Business, Enterprise and the API), OpenAI says it doesn’t train on your data by default.

Even with training off, you’re still sending customer data to a third party, so check your own privacy policy and any customer contracts. The simplest habit is still the best one: strip identifiers before you paste.

No handoff, no follow-up

A customer-facing assistant needs a way out: “I’m not sure, let me get a person.” Raw ChatGPT has no queue, no notification to your team and no record of which conversations need attention. The customer either gets a guess or gets stuck.

No analytics

When agents use ChatGPT in a browser tab, nothing about those conversations comes back to you. You don’t learn which questions keep coming up, which answers are missing from your docs, or which bug twenty people hit this week. The work gets done, but you learn nothing from it.

How to fix it: a docs-grounded support bot

The fixes for these problems are known. They just aren’t part of a general chat window. If you want AI to talk to customers directly, look for four things.

Retrieval from your own knowledge base

The bot should search your help center, docs, website and FAQ for each question and answer from what it finds, not from general knowledge. When the answer isn’t there, it should say so. This is what people mean by a knowledge base chatbot, and it deals with most of the “made-up policy” problem, because the model is working from your text instead of from what companies usually do.

Retrieval is only as good as the content behind it. If your refund policy page is out of date, the bot will quote the out-of-date version. Keep your docs current and read the unanswered questions regularly.

Handoff to a human

When the bot doesn’t know, or the customer asks for a person, the conversation should be flagged and someone on your team notified in the channel you actually watch. The customer should hear that a person is coming, not get a loop of “I didn’t understand that.”

Confirm before any action

Some questions need live data: order status, current plan, an invoice link. A support bot can call your API for these. For anything that changes something, like cancelling an order or updating an address, the bot should show what it’s about to do and wait for the customer to confirm. It should also know who it’s talking to, through a verified identity from your backend rather than whatever the visitor types.

Visibility into every conversation

You should be able to read any conversation, see which questions went unanswered and see which complaints repeat. That turns support from a cost you pay every day into a list of things to fix in the product and the docs.

ChatGPT for your agents vs a customer-facing bot

A short decision guide:

Most teams end up with both. The bot handles the repeated questions at any hour, and the people on the team use ChatGPT or a similar tool to write faster on the tickets that reach them. If you’re planning that split, our page on customer service automation covers what to automate first and what to keep for people.

Where Feedbot fits

Feedbot is one option for the customer-facing side. We build it, so weigh this section accordingly.

It’s an AI chat agent you add to your site with one script tag. It answers from your knowledge base, which you build from your website, documents or plain text and FAQ, and it replies in the visitor’s language even if your docs are in another one. When it doesn’t know, it says so, offers a person and logs the question as a gap. Handoff notifications go to Telegram, Slack, Discord or email.

On Pro and Business it can also call your HTTP API or functions on your page through Actions, for example to look up an order, and changes run only after the customer confirms. If you pass the signed-in user’s id with an HMAC hash from your server, the bot treats that identity as verified. See the AI customer service agent page for how that works.

Two things make it different from a plain support bot. The same agent qualifies leads when a visitor wants to buy. And bug reports and complaints are grouped into issues with user quotes and frequency, which a coding agent can pull through MCP or the npx feedbot pull CLI.

Pricing is flat: a free plan with 50 conversations a month, then $5, $15 or $39 a month, with no per-resolution fees and no per-seat pricing. The free plan uses knowledge you type in yourself; website import and document upload start on the $5 Starter plan. Details are on the pricing page.

If you only need help writing replies, you don’t need Feedbot. ChatGPT and the prompts above will do.

FAQ

Can ChatGPT replace a customer service team?

No. It can speed up the writing part of support and, inside a grounded bot, answer repeated questions from your docs. It can’t take responsibility for refunds, judgment calls or unhappy customers, and it doesn’t know your policies unless you give them to it.

Is it safe to paste customer tickets into ChatGPT?

Only after removing personal data, and only after checking your plan’s data settings. OpenAI states that individual accounts can have conversations used for model training unless that setting is turned off, while business offerings aren’t used for training by default. Your own privacy commitments to customers still apply either way.

How do I stop an AI support bot from making things up?

Ground it in your own content, tell it to say “I don’t know” when the answer isn’t there, give it a handoff path to a person, and review unanswered questions every week. Gaps in your docs are the most common cause of bad answers.

What’s the difference between ChatGPT customer service and a knowledge base chatbot?

ChatGPT answers from general training data and whatever you paste into the chat. A knowledge base chatbot searches your docs and site for every question, answers only from that, and runs on your website where customers can reach it directly.

Which ChatGPT plan should a support team use?

If agents will paste customer messages, look at the business plans, where OpenAI says data isn’t used for training by default, and confirm the current terms on OpenAI’s site before you decide.

Sources

• OpenAI Help Center, Data controls in ChatGPT: https://help.openai.com/en/articles/7730893-data-controls-faq
• OpenAI Help Center, How your data is used to improve model performance: https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance
• OpenAI, Enterprise privacy: https://openai.com/enterprise-privacy/
• OpenAI, Business data privacy, security and compliance: https://openai.com/business-data/