AI Prompts for Customer Service: 18 That Handle Real Tickets (2026)

The best AI prompts for customer service are not the ones that write pleasant emails. They are the ones that hold up on the hard tickets: the customer who is furious, the refund you have to refuse without losing the account, the bug you cannot reproduce, the reply you owe before you have an answer. The 18 prompts below are organized by those moments. Fifteen work in ChatGPT, Claude, or Gemini by pasting context in; the last three run on a connected inbox by themselves. Each comes with one line on why it works, because a prompt you understand is a prompt you can adapt.

One honest rule before any of them: the AI drafts, a human decides. Every prompt below asks for a draft or an analysis, never for something sent. And one privacy rule: customer emails contain names, addresses, and account details, so either strip them before pasting into a general chatbot or use a connection where data handling is explicit. The last section covers what that looks like.

Why hard-ticket prompts are different

A generic reply prompt produces a generic reply, which is fine until the ticket has stakes. The hard tickets fail on specifics: the apology that accidentally admits fault, the refusal that reads as a brush-off, the technical answer pitched at the wrong level. The prompts below all work the same way: they hand the AI the facts and the boundary (what you can offer, what you cannot, what tone the moment needs), and they ask for a draft that respects both. The facts are yours; the wording is the AI's job.

First replies and holding patterns (1-4)

1. The holding reply that does not sound canned. For when you owe a response before you have an answer.

Works in Claude, ChatGPT, or Gemini

A customer reported this issue: [paste]. We do not have a fix or an answer yet. Draft a first reply that tells them honestly what we know, what we are doing next, and when they will hear from us again, with a specific time. No corporate filler, no "we apologize for any inconvenience."

Copy

2. The information request without friction. Asking for details is where tickets stall; this asks for everything at once.

Works in Claude, ChatGPT, or Gemini

To diagnose this issue [paste the report], I need more information from the customer. List everything I plausibly need to ask, then draft one reply that asks for all of it in a numbered list, explains in one sentence why I need it, and thanks them without grovelling.

Copy

3. The expectation-setter. For requests that will take days, not hours.

Works in Claude, ChatGPT, or Gemini

A customer asked for [paste request]. Realistically this takes [X days] because [reason]. Draft a reply that commits to a date, explains the why in one sentence without jargon, and offers one useful thing they can do or have in the meantime.

Copy

4. The bad-news-first reply. Burying the answer is the most common service-writing failure.

Works in Claude, ChatGPT, or Gemini

Draft a reply to this customer [paste thread]. The answer to their request is no because [reason]. Put the no in the first two sentences, plainly and kindly. Then explain the why, then offer the nearest thing we can do: [alternative].

Copy

Angry customers and de-escalation (5-9)

5. The apology that does not grovel or admit fault. The two failure modes of service apologies, avoided in one prompt.

Works in Claude, ChatGPT, or Gemini

This customer is angry, with reason: [paste thread]. Draft a reply that apologizes specifically for what went wrong for them (not "any inconvenience"), takes ownership of the fix, and states exactly what happens next and when. Do not offer compensation yet, do not speculate about the cause, and keep it under 120 words.

Copy

6. The second-time complainer. The customer who has already complained once and is back angrier.

Works in Claude, ChatGPT, or Gemini

This customer complained before, we replied [paste our previous reply], and the problem happened again: [paste new complaint]. Draft a reply that acknowledges this is the second time without re-explaining the first, escalates the tone of ownership one level, and gives them a named point of contact and a concrete remedy.

Copy

7. The public-review threat. For "I will post this everywhere" messages.

Works in Claude, ChatGPT, or Gemini

A customer is threatening to leave negative reviews: [paste]. Draft a reply that does not mention the threat at all, addresses their actual problem with complete seriousness, and gives them a faster path than the one they were on. Calm, specific, zero defensiveness.

Copy

8. The wrong customer. Sometimes the customer is mistaken; saying so without condescension is a craft.

Works in Claude, ChatGPT, or Gemini

The customer believes [paste their claim], but what actually happened is [paste the facts, with evidence]. Draft a reply that walks through the facts gently, never uses the words "actually", "as previously stated", or "per our records", and leaves them a way to accept the correction without embarrassment.

Copy

9. The de-escalation triage. For deciding which angry threads need a phone call, not another email.

Works in Claude, ChatGPT, or Gemini

Here are the angry threads in our queue right now [paste]. For each one: rate the escalation risk from 1 to 5, say whether another email will help or make it worse, and flag any that should move to a call. One line of reasoning each.

Copy

Refunds, billing, and saying no to money questions (10-12)

10. The refund refusal that keeps the customer. The highest-stakes template in service.

Works in Claude, ChatGPT, or Gemini

A customer requested a refund: [paste request]. Our policy says no in this case because [reason], and I can offer [alternative: credit, discount, extension]. Draft a reply that leads with what I CAN do, states the policy limit in one plain sentence without quoting policy language, and reads like a person who looked for a way to help.

Copy

11. The billing-error apology. When the mistake is genuinely ours and money is involved.

Works in Claude, ChatGPT, or Gemini

We charged this customer incorrectly: [paste details: amount, what happened, when the refund lands]. Draft a reply that states the error plainly in the first sentence, gives the exact amount and refund date, explains in one sentence what we changed so it does not repeat, and does not pad any of it with mood words.

Copy

12. The cancellation save that respects the no. Retention flows that ignore the customer's stated reason lose twice.

Works in Claude, ChatGPT, or Gemini

A customer wants to cancel, reason given: [paste]. Draft a reply that takes their reason at face value, offers exactly one relevant alternative (not a menu), makes the cancellation path clear and easy in the same message, and thanks them either way. If their reason suggests we genuinely cannot serve them, say the cancellation is processed and leave the door open.

Copy

Troubleshooting and technical replies (13-15)

13. The step-by-step that matches the customer's level. Most troubleshooting replies are pitched wrong.

Works in Claude, ChatGPT, or Gemini

The fix for this customer's problem is: [paste internal notes or KB article]. Judging from how they wrote their message [paste it], match their technical level and draft numbered steps they can follow, with what they should see after each step. If any step could go wrong, say what "wrong" looks like and what to do.

Copy

14. The cannot-reproduce reply. "Works for us" is the most hated sentence in support; this replaces it.

Works in Claude, ChatGPT, or Gemini

We tried to reproduce this bug and could not: [paste what we tried]. Draft a reply that shows the customer exactly what we tested so they can spot the difference from their setup, asks for the two or three specifics most likely to explain the gap, and treats their report as true throughout.

Copy

15. The workaround-with-a-promise. For when the real fix is weeks away.

Works in Claude, ChatGPT, or Gemini

The proper fix for this issue ships in [timeframe]. The workaround until then: [paste]. Draft a reply that leads with the workaround in numbered steps, is honest that it is a workaround, commits to telling them when the real fix lands, and does not oversell either part.

Copy

The prompts that work the queue themselves (16-18)

Everything above waits for you to paste context in. These three fetch it themselves. They run in Claude or ChatGPT connected to a shared inbox through MCP, where the assistant has real tools with per-tool permissions instead of a pasted snapshot. Full disclosure: they are written against Drag's MCP server, which is our product; the tool names below are its real ones.

16. The morning service sweep. Runs the triage that section one approximates by hand.

Claude

Search our support inbox for every unanswered customer thread (search_threads), list them oldest first with one line each on what they need, and flag any where the customer's last message reads angry or mentions cancelling.

Copy

What happens: the assistant queries the connected inbox directly, so the list is live, not a paste from twenty minutes ago, and the flags land before the coffee does.

17. The history-informed reply. The reply that knows the customer's past without you digging for it.

Claude

Pull this customer's conversation history with us (get_contact_conversations), then draft a reply to their open thread that reflects anything we promised them before. Show me the draft; do not send it.

Copy

What happens: the assistant reads the relationship, not just the thread, and the draft stops contradicting last month's promise. Sending stays yours.

18. The Friday service report. The reporting prompt that ends the "how fast are we?" guesswork.

Claude

Using get_response_times and get_daily_activity for the last 7 days, give me: median first-reply time, the slowest three threads and why, and volume by day. Three bullets of what to fix next week.

Copy

What happens: the numbers come from the inbox's own analytics tools, so the Friday report writes itself from data instead of memory.

Setting this up takes minutes, not an integration project: the connect a shared inbox to Claude guide walks through it, and MCP for customer support covers what the protocol changes for support teams generally.

The privacy rule, stated properly

Customer emails carry names, order numbers, addresses, and sometimes payment details. Pasting them into a general chatbot puts that data wherever that chatbot's data policy puts it, which on some free tiers includes model training. The workable rules: strip identifying details before pasting when you use prompts 1 to 15; check the data-training setting on your AI plan (on Claude's individual plans it is called Model Improvement, in Privacy settings); and for anything running the queue itself, use a connection where the permissions are per-tool and explicit, which is what the MCP route above is for. A team that handles sensitive categories (health, finance, legal) should treat the paste-in prompts as off-limits for raw threads entirely.

Frequently asked questions

Do these prompts work in ChatGPT, Claude, and Gemini?

Prompts 1 to 15 work in any of the three: they carry their own context, so paste the thread and the boundary and any current model handles the rest. Prompts 16 to 18 need an assistant connected to an inbox through MCP, which today means Claude or ChatGPT connected to a tool that publishes an MCP server.

Is it safe to paste customer emails into ChatGPT?

Not raw, on a free tier: names, account details, and payment references go wherever the data policy puts them, which can include model training. Strip identifying details first, use a business plan where training is off, or use an MCP connection with explicit per-tool permissions instead of pasting at all.

Can AI handle an angry customer by itself?

It can draft the reply; it should not decide the outcome. The de-escalation prompts here (5 to 9) hand the AI the facts and the boundary, and they ask for a draft under 120 words that a human reads before anything sends. The judgment call, compensation, escalation, or a phone call, stays human.

How is this different from your shared inbox prompts page?

The shared inbox prompts page is about running a team's queue: triage, ownership, handoffs, one voice across many hands. This page is about the tickets themselves: what to write when the customer is angry, owed money, or stuck. Most support teams end up using both.

What is the best AI for customer service replies?

For drafting, any of the current assistants (ChatGPT, Claude, Gemini) writes a good reply when the prompt carries the facts and the boundary. The differentiation is connection, not prose: an assistant that can read the real thread history and the team's queue drafts from the truth. That is an MCP question, and our MCP for customer support guide maps which tools offer it.

Related guides: the rest of the prompt library

The shared inbox prompt library covers team-queue prompts: triage, handoffs, and team voice. For personal email, the ChatGPT prompts and Claude prompts guides go deeper on individual workflows. When you need ready-made human templates rather than prompts, customer service email examples has them. And for the setup that makes prompts 16 to 18 real: connect a shared inbox to Claude, or see how teams run customer support inside Claude or ChatGPT end to end.