Customer Service Reply Prompt for ChatGPT

Use this customer service reply prompt for ChatGPT when you need to answer a support message clearly without letting AI decide policy, expose customer details, or promise a fix you cannot give. The goal is to turn a messy support message into a clear reply that stays inside your real options and shows when a human should take over.

Best use cases

Replying to angry customers without sounding defensive.
Explaining refunds, delays, replacements, or missing information.
Classifying support tickets by request type, policy fit, urgency, and handoff need.
Making short support replies warmer and clearer.
Keeping simple requests from becoming long AI-sounding replies.
Turning a long complaint into the main issue and next step.
Handling AI-written complaints that sound formal but hide the customer's real request.
Checking whether your draft sounds robotic, cold, or like a generic template.

What to prepare

Before using ChatGPT, collect the customer's message, order or case status, support channel, ticket type, urgency, company policy, trusted FAQ or help-doc answer, related attachment or forwarded-thread context, and the action you can actually offer. Replace names, emails, phone numbers, addresses, order IDs, account details, and private attachment details with placeholders before pasting anything into an AI tool.

If the customer is quoting wrong AI information about your services, prices, hours, or policies, first audit the claim against your source-of-truth facts with the business information check, then write the support reply.

Copy this prompt

Write a customer service reply that is calm, useful, and policy-safe. Customer message: [paste message after removing names, emails, phone numbers, addresses, order IDs, and account details] Customer's exact question or request: [refund, replacement, delivery update, password help, cancellation, bug status, missing information, complaint, or "not clear"] Ticket type and urgency: [billing, shipping, account access, bug, how-to question, complaint, refund, cancellation, safety, legal, or unknown; low, normal, urgent, or unknown] Situation: [summarize order status, product, service, shipping, billing, bug, or delay] Attachments or messy context: [summarize attached document, screenshot, forwarded thread, long email chain, or "none"; paste only the relevant excerpt after removing private details] What we can offer: [refund, replacement, workaround, troubleshooting step, wait time, escalation, or "not sure yet"] Policy limits: [what we can and cannot promise] SLA or service commitment: [response time, support tier, contract rule, promised delivery window, regional rule, or "not applicable"] Trusted source: [FAQ, help-doc paragraph, order status, internal policy note, or "no trusted source yet"] Source conflicts or gaps: [old help article, unclear refund rule, missing order status, customer asks for a person, repeated failed bot answer, or "none known"] AI-written or legal-sounding claims: [customer cites policy, legal rights, contract language, fake case law, or "none"] Customer emotion: [angry, confused, disappointed, worried, urgent, or unknown] Human-owned opening: [the first 1 or 2 sentences a person wants to own, or "draft this carefully"] Brand voice: [plain, warm, direct, formal, casual, etc.] Channel and expected length: [email, chat, help desk note, social DM, short answer, full reply, or "match the customer's length"] Tone: Calm, helpful, concise, and human. Answer the customer's exact request first, then explain only what is needed. Do not sound defensive or like a generic AI template. Rules: - Do not invent policy, timelines, refunds, discounts, legal claims, or technical facts. - Use only the trusted source and situation details I provided. - If key information is missing, ask for it instead of guessing. - If the customer used AI to write a long or legal-sounding complaint, extract the exact unresolved request first. Do not answer fake legal claims or vague formal wording as if they are verified facts. - If no trusted source supports the answer, say what must be checked or hand it to a person. - If the trusted source conflicts with the order status, customer history, or another policy note, do not write a final answer. Mark the conflict and prepare a handoff summary. - If an attachment, forwarded thread, or long email chain is involved, separate what the customer provided from what your trusted source verifies. - If the customer asks for a person, says the bot already failed, or the same answer has not solved the problem, stop trying to resolve it with another long AI reply. - If the case needs judgment, SLA review, security review, legal review, medical advice, billing authority, account access, safety review, or a policy exception, mark it for human handoff. - For emotional, money-related, SLA-related, or high-value customer messages, treat AI as a draft or summary tool unless a person approves the final reply. - If this is a routine FAQ answer, say why it is safe to answer from the trusted source. - Keep private details out of the reply. - Match the channel and keep simple requests short. - Give one clear next step. Return: 1. Safe to answer now: yes, no, or needs review 2. A one-sentence direct answer to the customer's exact request 3. A short reply under 120 words, only if it is safe to answer 4. A warmer version 5. A firmer version 6. Source used for the answer 7. Source conflict or missing fact to verify 8. Attachment or messy-context summary, if any 9. Ticket classification: routine FAQ, policy-limited, sensitive, urgent, or needs investigation 10. Recommended workflow: saved FAQ reply, AI draft for review, human opening plus AI body, or full human review 11. Whether a human handoff is needed and why 12. If handoff is needed, a short handoff summary for the person taking over: customer issue, source checked, what was already offered, missing facts, conflict to resolve, and recommended next action 13. A one-sentence internal summary of the issue 14. AI-written complaint risk: the real request, any unsupported legal or policy claim, and whether to reply, call, or hand off 15. Any sentence that could sound defensive, robotic, too long, or too vague

Test input

Customer message: I paid for express shipping and the order is still not here. This is ridiculous. I want my money back. Customer's exact question or request: refund request Ticket type and urgency: shipping/refund, urgent Situation: Order shipped late because of a warehouse delay. Tracking now shows delivery tomorrow. Attachments or messy context: Customer attached a screenshot of the express shipping charge and tracking delay. No account document is pasted into ChatGPT. What we can offer: Refund express shipping fee, send tracking link, and escalate if it is not delivered tomorrow. Policy limits: We cannot refund the full order unless the package is lost or returned. SLA or service commitment: Express shipping was promised by Friday. Trusted source: Shipping policy says express shipping fees can be refunded when the carrier misses the promised window. Full order refunds require a lost package, return, or manager approval. Source conflicts or gaps: None known. Customer emotion: angry and urgent Human-owned opening: You're right to be frustrated that express shipping missed the promised window. Brand voice: plain, warm, direct Channel and expected length: email, short answer

What good output should include

A short acknowledgment that repeats the customer's real issue in plain words.
A direct answer to the customer's question before any explanation.
Only the remedy, refund, timeline, or escalation path you can actually offer.
One next step the customer can understand immediately.
A clear note about what source the answer used, such as the FAQ, help doc, order status, or policy.
A clear warning when the source conflicts with another fact or does not support a final answer.
A separate summary of attachments, screenshots, forwarded threads, or long email chains so the reply does not mix customer-provided context with verified facts.
A ticket classification that separates routine questions from sensitive, urgent, or policy-limited cases.
A workflow recommendation that says whether AI can draft the reply, a person should write the opening, or the whole case needs human review.
A human handoff note when the issue is complex, sensitive, or outside the written policy.
A short internal handoff summary so the next person does not make the customer repeat the whole story.
No blame, no long excuse, no fake empathy, and no over-apology.
No private customer details copied back into the reply.

Use it as a reply checker

For angry customers, write a rough reply first, then ask ChatGPT to check tone before you send it. This is useful when you are frustrated, rushed, or trying to be firm about a refund policy.

Act as a customer reading this support reply. Customer's likely emotion: [angry, confused, disappointed, worried, urgent, or unknown] My draft reply: [paste draft] Check: 1. Does the reply answer the customer's exact question first? 2. Does the reply clearly acknowledge the customer's issue? 3. Does anything sound defensive, robotic, dismissive, too long, or like a template? 4. Does it promise anything not stated in the policy? 5. Is there one clear next step? 6. Does the reply ignore a source conflict, missing fact, or direct request for a person? 7. Should this stay with AI drafting, use a saved FAQ answer, or go to a person? 8. Rewrite only the parts that need fixing.

Repair prompt

The reply still sounds wrong. Rewrite it again with these rules: - Keep the same facts, policy limits, and next step. - Start with the direct answer, then acknowledge the customer's problem. - Remove corporate filler and over-apology. - Use shorter sentences. - If we cannot solve it yet, say what we will check next and when we will follow up. - Match the support channel and do not make a simple answer longer than needed. - Do not add any new promise, refund, discount, or timeline.

When to stop and escalate

Do not use a prompt as the final authority for legal threats, chargebacks, medical or safety issues, harassment, account security, or anything involving regulated personal data. Use ChatGPT to organize the facts, then send the case to the right person.

Use AI freely for repetitive questions when the answer is already in your FAQ, order status, or written policy. Keep a person in the loop when the customer is angry, the facts conflict, the answer affects money or access, the customer asks for a person, or the reply would require an exception.

AI-to-human handoff prompt

If the customer asks for a person, the bot has already failed, or the answer needs account, billing, policy, safety, or legal judgment, use ChatGPT to prepare the handoff instead of another customer-facing reply.

Prepare this support case for human handoff. Do not write another final customer reply. Customer message: [paste anonymized message] What the AI or support team already tried: [summarize prior answers, links, troubleshooting steps, or "none"] Trusted source checked: [FAQ, policy, order status, account note, bug report, or "no trusted source yet"] Reason for handoff: [customer asked for a person, repeated failed bot answer, billing/refund, account access, legal/safety, source conflict, missing authority, or unknown] Return: 1. One-sentence issue summary 2. Customer's exact unresolved request 3. What was already tried 4. Source checked and what it supports 5. Missing facts the human must verify 6. Risk level: low, medium, or high 7. Best owner: support agent, billing, technical support, manager, legal/safety, or unknown 8. Suggested next action for the human 9. Short customer update that says the case is being handed to the right person without promising an outcome or exact timeline Rules: - Do not make the customer repeat information already provided. - Do not invent account status, refund eligibility, policy exceptions, technical causes, or timelines. - Keep private details out of the summary unless the receiving human needs them in the real support system.

Common mistake

Do not ask ChatGPT to decide the policy. Give it the policy, available options, missing facts, exact customer request, and channel, then ask it to write or check the message. A simple customer question should not become a long AI support script.

When a customer reply suddenly becomes long, formal, or legal-sounding, do not let AI write another long counterargument. Ask it to identify the real unresolved request, unsupported claims, verified source, and best human owner. Often the useful answer is a short clarification, a phone call, or a handoff, not another polished paragraph.

Specific customer service prompts

Find repeated ticket problems

If you have several similar tickets, use ChatGPT to find patterns before rewriting the same answer again. Remove private details first, then ask for common issue types, missing FAQ answers, and where a person should step in.

Analyze these support tickets for repeat problems. Do not write customer replies yet. Ticket samples: [paste 5 to 20 anonymized ticket summaries] Trusted sources: [FAQ links, help-doc excerpts, policy notes, or "not available"] Return: 1. Common ticket categories 2. Questions that can be answered from the trusted source 3. Questions that need a better FAQ or help-doc answer 4. Tickets that need human review and why 5. Missing facts agents should collect before replying 6. One short draft macro or saved reply for each safe routine category 7. Risks: privacy, refunds, account access, legal, safety, or policy exceptions Rules: - Do not invent answers for tickets that lack a trusted source. - Do not include private customer details in the macros. - Mark anything unclear as "needs review."

Sort a busy support queue first

When your inbox has many similar questions, do not ask ChatGPT to answer everything at once. Use it first to separate routine tickets from cases that need a person, missing information, or a source check.

Triage these support tickets before anyone writes replies. Ticket list: [paste anonymized ticket summaries only] Trusted sources available: [FAQ, policy page, help-doc excerpt, order-status field, saved macro, or "none yet"] Return a table with: 1. Ticket number 2. Customer's main request 3. Routine answer, missing information, source conflict, urgent issue, or human handoff 4. Trusted source needed before replying 5. One fact an agent must verify 6. Suggested owner: support agent, billing, technical support, manager, legal/safety, or unknown 7. First next step Rules: - Do not write final customer replies yet. - Do not guess policy, refund eligibility, timelines, account status, or technical cause. - Put angry customers, repeated failed bot answers, account access, money, safety, legal, or policy exceptions into human handoff.