AI prompts for customer service with a check before sending
AI prompts for customer service teams: 20 copy-ready prompts for ChatGPT, Claude, Gemini or Microsoft Copilot to draft replies, summarize threads, write help articles and check quality. Each says what to paste and what to check before sending.
Made by Anatolii Ulitovskyi 
Fact-checked by RH Fardin Updated 1. Customer service AI prompts: the 5-part prompt
Start hereThese customer service AI prompts work in any general AI assistant: ChatGPT, Claude, Gemini or Microsoft Copilot. Google's prompting guide names four parts of a prompt: persona, task, context and format1. We add a fifth, rules, to stop invented refunds and dates. You decide; a person checks and sends.
The 5-part prompt (fill in once, reuse)ROLE: You are a customer support agent at business, a type of business. Write in language, in a warm and plain tone, signed your first name.
TASK: what to write, for whom.
FACTS (use only these): the customer's message with the first name only, the facts, the policy line, my decision.
RULES: Add nothing that is not in the facts. No prices, dates, refunds, discounts or promises unless they are in the facts. If a fact is missing, write MISSING and tell me what you need.
FORMAT: email or chat, under 80 words, one next step at the end.
Never paste into an AI tool
- Card numbers, bank details, passwords or one-time codes.
- A surname with an email, phone number or home address. Use the first name, or Customer A.
- Health details, ID numbers, other customers' messages, internal notes about a person.
Check before sending (every prompt)
- Facts: each one is in the order, the ticket or the policy.
- Numbers: dates, amounts and times copied exactly.
- Promises: no refund, credit, discount or date you did not decide.
- Tone: no blame, no sarcasm, one apology at most.
- Privacy: nothing about other customers, no internal notes.
Why a person checksAI can give wrong answers that sound confident; OpenAI advises using ChatGPT as a first draft, not a final source2. And the business answers for what it sends: a Canadian tribunal held an airline responsible for wrong information from its chatbot3.
For your website AI agent, see the AI agent persona guide and the chatbot fallback message examples. For sales work, see the sales AI prompt templates.
2. Draft a customer reply from the facts and your decision
Most usedPaste the facts and your decision, not just the question. The AI only writes the words.
2.1Reply from the facts
Paste inThe message (first name only), the facts, your decision.
Check before sendingDates and amounts match the order. Nothing you did not decide.
PromptYou are a support agent at business. Write a reply to this customer message:
customer message
Facts (use only these): order or account facts, dates, status
Our decision: what we will do and when
Answer the question in the first sentence. Apologize once, for the specific problem. Add no dates, amounts, refunds, credits or discounts that are not above. If a fact is missing, write MISSING.
Email or chat, under 80 words, one next step at the end.
2.2Say no, from the policy
Paste inThe request, the policy text, your decision, one real alternative.
Check before sendingThe no comes first. The policy is quoted right. No exception you cannot give.
PromptWrite a reply to a customer who asked for request. Our answer is no.
Policy (quote only this): policy text
What we can offer instead: alternative, or nothing
Give the decision in the first 2 sentences, explain the one policy reason in plain words, then offer the alternative once. Do not apologize for the policy, hint that it may change or invent exceptions. Under 100 words, warm and plain tone.
2.3Ask for one missing detail
Paste inThe message and the one detail you need.
Check before sendingOne question. It never asks for card numbers, passwords or ID numbers.
PromptWrite a short reply to this message:
customer message
We need one detail before we can help: detail, such as the order number. Ask for it in one question, say why we need it and what happens next: next step and time. Never ask for card numbers, passwords or ID numbers. Under 40 words.
3. Fix the tone, shorten or translate a support reply
Write the facts yourself, then let the AI fix how they sound. Your facts stay word for word.
3.1Calm the tone
Paste inThe customer's message and your draft.
Check before sendingCompare every date and amount with your draft. No new promise.
PromptRewrite my reply to an upset customer.
Their message: customer message
My draft: your draft
Keep every fact, date and amount exactly. Name the problem in their words first. Take responsibility for our part. Remove defensive phrases such as "as per our policy". End with what happens next and when. Same length or shorter.
3.2Shorten and simplify
Paste inYour reply and the facts that must stay.
Check before sendingThe facts you listed and the link are still there and correct.
PromptShorten this reply to under 60 words for chat or a phone screen:
your reply
Keep these exactly: facts that must stay. Use short sentences and everyday words, no jargon, at most one apology. Return only the new reply.
3.3Translate and back-translate
Paste inYour checked reply.
Check before sendingThe back-translation matches. Dates are in the local format. A fluent speaker checks policy wording.
PromptTranslate this reply into language for a customer in country:
your reply
Keep names, order numbers, amounts, dates and links exactly as written. Use the formal or informal form of address. Then translate your version back into English, sentence by sentence, so I can check it.
In CRMsoftware.proOn Growth, Auto translate in Live Support translates chats for your team, and the website AI agent replies in 95+ languages. Use 3.3 for emails and help articles.
4. Summarize a long support thread for handover
Use these before you hand a ticket over or ask a manager to decide. Replace names with Customer and Agent first.
4.1Handover summary in 5 lines
Paste inThe whole thread, with names replaced.
Check before handing overEach promise under Done so far is really in the thread, with its date.
PromptSummarize this support thread for a colleague who has not seen it:
thread
Use exactly these 5 lines:
Wants: what the customer wants, in one sentence.
Facts: products, order dates and amounts mentioned.
Done so far: what we tried or promised, with dates.
Open: what is still unanswered.
Next: the one next step, who owes it and by when.
Use only the thread. Write UNCLEAR where it is not clear.
4.2Ask a manager to decide
Paste inThe 5-line summary, the policy and your options.
Check before sendingAmounts and the customer's deadline are right. No personal data beyond a first name.
PromptWrite a short internal message asking manager to decide on a support case.
Summary: 5-line summary
Customer asks for: request
Policy: policy text
Options: 2 or 3 options with their cost
My recommendation: option and why
Answer needed by: date and time
Under 100 words, options as a numbered list. End with the question to decide.
5. Turn solved tickets into help articles and saved replies
When the same question arrives 3 times, write the answer once and reuse it.
5.1Help article from solved chats
Paste in3 to 5 solved chats with names removed, and the answer that worked.
Check before publishingFollow every step yourself. Labels and screenshots match today's product.
PromptWrite a help article about topic from these solved chats:
3 to 5 chats
The answer that worked: steps
Title: the question as a customer would type it. Then a one-sentence answer, numbered steps with one action each, what the customer sees when it worked, and when to contact us. Use only the answer that worked. Then list questions from the chats the article does not answer.
5.2FAQ entry for the knowledge base
Paste inThe question in the customer's words and the approved answer.
Check before publishingIt matches your policy page. Prices and dates are current or left out.
PromptWrite one FAQ entry.
Question, in the customer's words: question
Approved answer: answer
Condition or exception, if any: condition
Start the answer with yes, no or the number. At most 3 sentences. Use only the approved answer.
5.3Saved reply with placeholders
Paste inA reply you checked and sent.
Check before savingNo real name, date or amount is left in. The reply follows today's policy.
PromptTurn this reply into a reusable saved reply for topic:
reply you sent
Replace every detail that changes per customer, such as names, order numbers, dates and amounts, with a short placeholder in square brackets. Keep everything else word for word. Suggest a title and a lowercase shortcut with no spaces that starts with a slash.
Where they goHelp articles and FAQ entries go into your website AI agent's Knowledge Base (see the FAQ knowledge base starter); saved replies into Settings > Saved Replies on Growth.
6. Check a reply against the policy, find contact reasons
For team leadsFor the team lead: a second pair of eyes before an important reply, and a weekly look at why customers write.
6.1QA check against the policy
Paste inThe customer's message, the draft, the facts and the policy.
Check before sendingYou still read the draft yourself: a reviewer AI can miss things too.
PromptAct as a strict quality reviewer. Check this draft reply before it is sent; do not rewrite it.
Customer message: customer message
Draft: draft reply
Facts and policy: facts and policy text
Make a table: each claim in the draft (dates, amounts, promises, product facts), Supported (Yes, No or Not in the facts) and the fix. Then flag blame or sarcasm, a missing next step, and personal data that should not be there.
6.2Top contact reasons
Paste inLast week's first messages, names and order numbers removed. Survey comments and reviews work too.
Check before actingRecount the top 2 reasons by hand. AI counts can be off.
PromptHere are number customer messages from last week, one per line, names and order numbers removed:
messages
Group them into at most 8 reasons. For each: a short name, the count, one example in the customer's words, and what would prevent it: a help article, a saved reply, or a fix in the product, shop or policy. Count each message once. Sort by count.
Next weekTurn the top reasons into tags in Settings > Conversation Tags (New tag) to categorize next week's chats.
7. AI prompts for ecommerce and SaaS customer service
Shops get order and return questions; software companies get how-to and bug questions. Copy the facts from the source; never let the AI guess.
7.1Order status reply (ecommerce)
Paste inOrder date, shipping status and the carrier's estimate, copied today.
Check before sendingThe status matches the carrier page right now. No "it arrives tomorrow" unless the carrier says so.
PromptWrite a reply to a customer asking where their order is.
Facts: ordered date, shipped date or not yet, carrier status status and date, estimated delivery date or none.
Our next step: what we will do and when.
Do not guess a delivery date. Under 70 words, chat or email.
7.2Return or refund reply (ecommerce)
Paste inThe request, the policy and what you approved in your shop or payment platform.
Check before sendingAmount, method and timing match what you approved. The refund is made in that platform.
PromptWrite a reply about a return.
Request: request
Return policy: policy text
Decision: approved or not, amount, refund method, timing
What the customer does next: label, packing, drop-off, deadline
Give the decision first. State the amount and timing exactly as above. Under 100 words.
7.3Bug report for developers (SaaS)
Paste inThe conversation with names and emails removed, plus the plan and browser.
Check before sendingReproduce it yourself. No screenshot that shows customer data.
PromptTurn this support conversation into a bug report for our developers:
conversation
Write: a title of up to 10 words, steps to reproduce, expected result, actual result, the plan, device and browser, how many customers reported it, and the workaround we gave. Write UNKNOWN for anything the conversation does not say. Do not guess the cause.
Businesses that take bookings: see the appointment booking AI prompt templates. Insurance: see the insurance AI prompt templates.
8. AI prompts for customer onboarding and retention
One goal per message: a first step for new customers, a reason to come back for quiet ones, a clear way out for those who leave.
8.1Welcome message
Paste inWhat they chose, the first step that helps most, where to get help.
Check before sendingThe link works and the step fits their plan. No promised results.
PromptWrite a welcome email or message from business to a new customer who just signed up for or bought what.
Facts: what they chose. First step that helps most: first step. Help: link, or reply to this email.
Thank them in one sentence. One goal: get them to take the first step. No feature list, no discount. Under 90 words.
8.2Check-in when a customer stalls
Paste inWhat they did so far and where they stopped.
Check before sendingCheck the account first: they really stalled. Sent by a person.
PromptWrite a short check-in to a customer who stopped at which step, or has not ordered or logged in since when.
What they did so far: facts
Likely blocker: blocker
Offer one thing that removes it: guide, setup help or a 15-minute call.
Helpful, not pushy. Never suggest they did something wrong. Under 70 words, one question at the end.
8.3Reply to a cancellation request
Paste inTheir reason, how to cancel, the one alternative you offer.
Check before sendingThe cancel steps work. The offer is approved, and made once.
PromptWrite a reply to a customer who wants to cancel plan, subscription or order.
Their reason: reason, in their words
How to cancel: steps or link
What happens after: access, data, last charge
One alternative we can offer: pause, smaller plan, fix, or nothing
Confirm how to cancel first, then offer the alternative once. No guilt, no pressure. Under 90 words.
Win-back and feedbackFor survey comments, NPS answers or reviews, use 6.2. For a win-back message, use 8.2 with the reason they left as the blocker.
For small sales and support teams
The AI agent answers visitors. Your team sends the checked replies.
In CRMsoftware.pro the website AI agent answers from your Instructions and Knowledge Base in 95+ languages and hands chats to your team in Live Support. Save the replies that pass your checks as Saved Replies.
- AI chat 24/7 from your Instructions and Knowledge Base
- Human handover in Live Support from Launch
- Saved Replies and inbox auto-translation on Growth, $49 a month billed yearly
Free plan: 3 users, no credit card.
Filled example
One damaged order: the prompt, the 5 checks, the reply sent
Quillfeather Stationery, a fictional online shop for notebooks and pens, keeps its prompts in the Excel library. On Monday 12 October 2026, Rohan Desai wrote in the website chat about order QF-20417. Noelle Kapoor, the support lead, took over, drafted her reply with prompt 2.1, ran the 5 checks and sent it.
Prompt 2.1, as Noelle filled it in
| Prompt | 2.1 Reply from the facts, version 2 |
| AI assistant | The shop's work account |
| Pasted | Rohan's message (first name only), the order contents, the policy line, Noelle's decision, the carrier's date |
| Left out | Surname, email, address, order number, the photo |
| Decision (Noelle's) | New pen and notebook by express, free, no return (policy: damage reported within 30 days) |
The 5 checks before sending
| Facts | Pen and notebook, express, free: as decided |
| Numbers | Wednesday 14 October: matches the carrier's estimate |
| Promises | Deleted "we've also refunded your shipping": not decided. Logged as a Fail, fixed in version 3 |
| Tone | 66 words, one apology, no blame, one next step |
| Privacy | No other customer, no internal notes |
What lands in the app and the CRM
| Live Support | Awaiting agent 9:44, claimed by Noelle 9:58, resolved 10:12 |
| Notes | On Rohan's contact: QF-20417, pen leaked, notebook stained; replacement by express on 12 October, no return |
| Tasks | Check QF-20417 tracking; Assignee: Noelle; Deadline: Wednesday 14 October |
| Saved Replies | /damaged-item: the checked reply with brackets, shared with the team |
Test log, prompt 2.1
| Version | Last test | Result |
|---|---|---|
| Version 3 | 13 October | 3 of 3 passed |
Why it matters
Why AI prompts for customer support need a person to check
In a study of 5,179 customer support agents, access to an AI assistant raised the number of issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with minimal impact on the most experienced and skilled4.
The risk is the same text sent unchecked. OpenAI says ChatGPT can produce incorrect or misleading outputs, sometimes sounding confident even when wrong2. And the business answers for what it sends: a Canadian tribunal held an airline responsible for all the information on its website, including what its chatbot told a customer3.
So every prompt here gives the AI only the facts, keeps the decision with you and ends with what to check before sending.
14%
more issues resolved per hour with an AI assistant, in a study of 5,179 support agents
34%
for novice and low-skilled agents in the same study
5 checks
before any AI draft reaches a customer (this template's rule)
Data per tool
Customer data in ChatGPT, Claude, Gemini and Copilot
Before anyone pastes a customer message, check which account they use and what its data settings say.
| AI assistant | Personal account | Work account |
|---|---|---|
| ChatGPT | Settings > Data controls: with Improve the model for everyone off, new chats are not used to train models. Temporary Chat is not used either5. | OpenAI does not use ChatGPT Business, Enterprise or Edu content to train its models by default5. |
| Claude | Privacy settings: the Model Improvement setting decides whether your chats help improve Claude. Incognito chats are not used6. | Anthropic does not use inputs or outputs from Claude for Work or its API for training by default7. |
| Gemini | Keep Activity off: future chats are not used to train Google's models, but are kept for 72 hours. Google asks you not to enter confidential information you would not want a reviewer to see8. | Work or school accounts may have different data terms: ask your admin8. |
| Microsoft Copilot | Use the work account your company gives you for customer messages, not a personal one. | Prompts, responses and data accessed through Microsoft Graph are not used to train foundation models9. |
Settings change: checked on 14 October 2026. Check again before you paste customer messages, and follow your company's AI policy.
Never paste
What never goes into a support prompt, and AI disclosure
Six rules for the whole team.
Payment and login details
Card numbers, bank details, passwords and one-time codes never go into an AI tool. No reply needs them.
Paste insteadPaid by card on date
Names with contact details
Personal data is any information about an identified or identifiable living person, such as a name, a home address or an email address10. Use the first name, or Customer A.
Paste insteadfirst name wrote: message
Only what the reply needs
Under the GDPR, personal data must be limited to what is necessary for the purpose11. Paste the facts for this reply, not the account history.
Other people's data
No other customers' messages, internal notes about a customer or staff complaints: a summary repeats whatever you paste.
AI disclosure for chatbots
Under the EU AI Act, AI systems that interact directly with people, such as a website AI agent, must be designed so that people are told they are interacting with an AI, unless that is obvious, at the latest at the first interaction12. This applies from 2 August 202613.
AI disclosure for drafts a person sends
Article 50 does not name replies a person writes with AI help, checks and sends. Its rule for AI-generated text covers text published to inform the public on matters of public interest, with an exception for human review and editorial responsibility12. Follow your own policy.
This is not legal advice. Check the rules for your country and your industry, and your company's AI policy.
Excel library
Customer service prompt library and test log in Excel
Four sheets: Library, Test log, Summary and Settings. Auto columns fill in by themselves. Opens in Excel, LibreOffice and Google Sheets.
| Column | What it does |
|---|---|
| Library: #, prompt, task group, version, owner | All 20 prompts from the tabs. Raise the version every time you change a prompt. |
| Prompt text, paste in, check before sending | The prompt with its brackets, what to paste and what to check. |
| Tests, passed, failed, last tested Auto | Counted from the Test log, for the current version only. |
| Retest by Auto | Last test plus 90 days (Settings); red when the date has passed. |
| Status Auto | Ready (3 passes, no fail), Test more, Fix and retest or Not tested. |
| Test log: date, prompt, version, test input | One row per test, on a real case with customer data removed. |
| Customer data removed, expected, result, checked by | Yes or No, what a pass looks like, Pass or Fail, who checked. |
| Counts for Auto | Current version, Old version or Not in library. |
| Summary Auto | Prompts by status, tests passed and failed, the pass rate, tests with customer data left in. |
In the sample, Quillfeather has tested 5 of the 20 prompts: 2 are Ready, 2 need more tests, 1 needs a fix. 11 of 15 tests passed (73%). Prompt 2.1 took 3 versions: version 1 invented a discount code, version 2 a shipping refund.
How to use
How to use these customer service AI prompt templates
- 1
Pick the prompt
Find the task in tabs 2 to 8, or start with 2.1.
- 2
Fill in the brackets
Business, tone and word limit once; fresh facts each time.
- 3
Decide, then paste
You decide refunds, dates and exceptions. The AI only writes.
- 4
Check, then send
Run the 5 checks from tab 1. A person sends every reply.
- 5
Log and version
Log each test in Excel. Any change to a prompt is a new version, tested 3 times.
Set it up in CRMsoftware.pro
- In My Bots, choose the bot. The website AI agent answers from Instructions and the Knowledge Base on the Basics tab; write those with the AI agent persona guide.
- Add help articles and FAQ entries from tab 5 to the Knowledge Base, click Save Bot and ask the question in Try Bot.
- When a chat is Awaiting agent in Live Support, click Claim. Draft with a prompt, run the 5 checks, send, then click Resolve.
- On Growth: Settings > Saved Replies > New reply. Add a Title, a Shortcut such as /damaged-item and the Reply body, then turn on Share with team. Type the shortcut in the chat reply box to insert it.
- Put the facts and your decision in Notes on the contact, and any follow-up in Tasks > New Task with an Assignee and a Deadline.
AI chat 24/7 and a basic knowledge base are on the Free plan: 3 users, no credit card. Human handover to Live Support comes with Launch, $19 a month billed yearly; Saved Replies and auto-translation in the inbox with Growth, $49 a month billed yearly.
A prompt is ready when
- 3 passed tests at the current version.
- Every bracket says what to fill in.
- Its check is in the library.
- Nobody pastes card numbers or full contact details.
FAQ
Customer service AI prompts: common questions
What are the best AI prompts for customer service?
The ones that paste the facts and your decision, set rules and say what to check. Start with 2.1 Reply from the facts: it adds no dates, refunds or discounts that are not in the facts. For sales work, see the sales AI prompt templates.
How do I write customer service prompts for ChatGPT, Claude or Gemini?
Use 5 parts: role, task, facts, rules and format. Google's prompting guide names persona, task, context and format1; the rules part stops invented promises. Fill in the business, tone and word limit once, and paste fresh facts each time.
Can I paste customer data into an AI assistant?
Only what the reply needs, and never card numbers, passwords or full contact details: a name with an email or home address is personal data10. Check your account's data settings first, such as Settings > Data controls in ChatGPT5; work accounts have their own terms.
Do I have to tell customers that AI helped write a reply?
For a chatbot in the EU, yes: AI systems that interact directly with people must tell them so unless it is obvious12, from 2 August 202613. Article 50 does not name a reply that a person checks and sends, so follow your company's policy. This is not legal advice.
Can AI summarize a support ticket for handover?
Yes. Prompt 4.1 turns a thread into 5 lines: wants, facts, done so far, open, next. Replace names first, and check that every promise in the summary is in the thread.
What are good AI prompts for ecommerce customer service?
Tab 7 has order status and return or refund replies; tabs 2 and 3 fit any shop. Paste the carrier status and the approved refund exactly; never let the AI guess a delivery date.
Which AI prompts help with SaaS support, onboarding and retention?
For SaaS: 7.3 for bug reports, 5.1 for help articles. For onboarding and retention: 8.1 welcome, 8.2 check-in when a customer stalls, 8.3 cancellation reply. One goal per message, and no offer you have not approved.
How do I test and version customer service prompts?
Run each prompt on 3 real cases with customer data removed and log them in the Excel test log. Any change makes a new version, and only tests of the current version count: Ready means 3 passes and no fail.
How do these prompts work with CRMsoftware.pro?
Run them in your own AI assistant. In CRMsoftware.pro the website AI agent answers visitors from your Instructions and Knowledge Base in 95+ languages and hands chats to Live Support from Launch, $19 a month billed yearly14. On Growth, $49, save checked replies as Saved Replies14.
Related templates
Free tools
Sources
- Prompt structure, April 2024: Gemini for Google Workspace: Prompting guide 101
- OpenAI Help Center: Does ChatGPT tell the truth?
- Civil Resolution Tribunal of British Columbia, 14 February 2024: Moffatt v. Air Canada, 2024 BCCRT 149
- Study of 5,179 support agents, NBER Working Paper 31161, published in The Quarterly Journal of Economics, 2025: Brynjolfsson, Li and Raymond, Generative AI at Work
- OpenAI Help Center, read 14 October 2026: Data controls in ChatGPT
- Anthropic Privacy Center, consumer plans: Is my data used for model training? (Free, Pro and Max)
- Anthropic Privacy Center, commercial products: Is my data used for model training? (Claude for Work and API)
- Gemini Apps Help, updated 13 October 2026: Gemini Apps Privacy Hub
- Microsoft Learn, updated 18 August 2026: Data, Privacy, and Security for Microsoft Copilot
- European Commission, personal data: Data protection explained
- European Commission, data minimisation: How much data can be collected?
- EU AI Act, Article 50, text as at 27 July 2026: AI Act Service Desk, Article 50: Transparency obligations for providers and deployers of certain AI systems
- EU AI Act, date of application: AI Act Service Desk, Article 113: Entry into force and application
- Plans and prices, checked 14 October 2026: CRMsoftware.pro pricing
Prompt in your AI assistant. Keep the checked answer.
Start free with 3 users and AI chat 24/7. Launch, $19 a month billed yearly, adds human handover in Live Support; Growth, $49, adds Saved Replies and inbox auto-translation.