The short answer
The best prompts for customer support give the AI your policy, your tone guide and an anonymised customer message, then ask for a reply that acknowledges the problem first and promises nothing outside the policy. Use them for angry customers, refusals, help articles, saved replies, ticket themes and incident updates. Claude is our pick for replies and ChatGPT for ticket analysis. Never paste customer names or order details into a personal plan; anonymise first or use an approved work plan.
How to use these customer support prompts
Customer support is one of the jobs where AI saves the most time, because so much of the work is writing: replies, saved replies, help articles, escalations and reports. These 12 prompts are for support agents, team leads and founders who answer customers themselves. They follow one principle: the AI writes, and your policy decides. Every reply prompt tells the AI what you can and cannot offer, so it never promises a refund or a fix you cannot deliver.
Replace everything in [square brackets]. Paste your actual policy, product names and tone guide. Without them, an AI reply sounds polite but generic, and customers notice. For the bigger picture, see how to use AI for customer service and AI for customer support teams.
Keep customer data out of consumer plans
Support tickets are full of names, emails, addresses, order numbers and sometimes payment or health details. Do not paste them into a free or personal AI plan. On the personal plans of ChatGPT, Claude and Perplexity, chats can be used for training unless the user switches it off, and that is not a decision a support agent can make for a customer. Anonymise first, by replacing names and numbers with placeholders such as [CUSTOMER] and [ORDER], or use a work plan your company has approved. ChatGPT Business and Enterprise and Claude Team and Enterprise do not train on your content by default, and Microsoft Copilot with a work account runs under enterprise data protection. Our guide to private AI for sensitive data covers the options.
Prompts for replying to customers
Calm an angry customer, Say no without losing the customer, Explain a technical fix step by step and Rewrite a reply in our brand voice cover most of a support inbox. The angry customer prompt puts the acknowledgement in the first sentence, because a reply that opens with the policy feels like a wall. The refusal prompt bans the phrase customers hate most and makes the no clear early, which is kinder than a long paragraph that hides it. Always read a reply before sending: you know the customer's history, and the AI does not.
Prompts for your help centre and saved replies
Turn a solved ticket into a help article and Build a set of saved replies turn work your team has already done into self-service. Every good answer an agent writes once is a help article waiting to happen. Phrase article titles as the question customers actually type, since that is how they search your help centre and how AI answer engines find it.
Prompts for finding what customers struggle with
Find the themes in a ticket export and Write the weekly voice of the customer report turn a pile of tickets into decisions. Export only the fields you need, remove personal data, and check the AI's counts against a sample, because AI tools can miscount or merge themes in long lists. Claude and ChatGPT both read attached files, so a CSV export works better than pasting.
Prompts for escalations, incidents and quality
Write an incident update, Escalate a ticket clearly, Score a reply against our quality checklist and Train a new agent with practice tickets help team leads keep standards high. The incident prompt tells the AI not to guess at causes, because a wrong explanation on a status page is hard to take back. The quality prompt only works with your own checklist, so paste it in full and keep scoring consistent across the team.
How to adapt these prompts
- Add the channel. Email, live chat, social media and phone scripts need different lengths and tone. Name the channel in every reply prompt.
- Save your context once. Put your tone guide, policies and product names in a project or custom instructions on your approved tool, so you do not paste them every time.
- Translate with care. AI can draft replies in the customer's language, but have a fluent speaker check saved replies before they go live.
- Keep a human on edge cases. Complaints that mention legal action, safety, vulnerable customers or cancellations of large accounts go to a person.
Which AI tool works best for customer support
For drafting replies and help articles, Claude is our pick, because its writing sounds human and it keeps facts and commitments exactly as you gave them when you ask it to. ChatGPT is strong for analysing ticket exports and building reports; free users get GPT-6 Luna, and Business, for 2 to 200 people, adds admin controls and no training on business data by default. Teams on Microsoft 365 can use Microsoft Copilot in Outlook and Teams under enterprise data protection. Grammarly helps agents write cleaner replies in the browser, and its Business suite plan adds a Zendesk connector for its Go assistant. See the best AI for customer support for a full comparison.
Prompts to copy
Calm an angry customer
Draft a reply to this customer message: [paste, with name, email and order details removed]. They are angry about [issue]. Our policy allows: [what we can and cannot do]. Acknowledge the specific problem in the first sentence, apologise once without blaming anyone, explain what we will do and by when, and end with one clear next step. Under 150 words, warm and plain. Do not promise anything outside the policy.
You get: A calm, policy-safe reply that addresses the real complaint
Say no without losing the customer
A customer is asking for [refund / exception / feature] that we cannot give because [reason]. Their message: [paste, anonymised]. Write a reply that says no clearly in the first two sentences, explains the reason in plain words, offers the best alternative we do have: [paste] and leaves the door open. Under 140 words. No corporate phrases such as "we apologise for any inconvenience".
You get: A clear no with a real alternative
Explain a technical fix step by step
Turn these internal troubleshooting notes: [paste] into instructions for a non-technical customer using [product] on [device or platform]. Numbered steps, one action per step, the exact button or menu names from our notes, and what they should see after each step. Add a final line on what to send us if it still does not work.
You get: Customer-ready steps from engineer notes
Rewrite a reply in our brand voice
Here is our support tone guide: [paste, or describe it in three adjectives with an example]. Rewrite this draft reply to match it: [paste]. Keep every fact and commitment exactly the same. Then tell me which sentences you changed and why.
You get: A reply that sounds like your brand, with the facts untouched
Turn a solved ticket into a help article
Here is a solved support conversation, anonymised: [paste]. Turn it into a help centre article: a title phrased as the question customers ask, a one-sentence answer, numbered steps, common mistakes and when to contact support. Write for someone who is in a hurry. Flag any step that depends on the customer's plan or settings.
You get: A help article from work your team already did
Build a set of saved replies
Our ten most common ticket types are: [paste]. For each, write a saved reply (macro) under 100 words with [placeholders] for the customer's name and specific details, a version for chat that is under 40 words and a note for agents on when not to use it. Keep the tone [describe].
You get: Ready macros for email and chat, with usage notes
Find the themes in a ticket export
Here is an export of [number] recent tickets with personal data removed (subject, category and the first message only): [paste or attach]. Group them into themes, count the tickets in each, quote two typical examples per theme and rank the themes by volume. Show how many tickets you could not place in a theme, so I can check the counts add up. Then suggest which three could be reduced with a product fix, a help article or a better saved reply.
You get: A ranked list of what customers contact you about, and what to fix first
Write the weekly voice of the customer report
Using these theme counts and example quotes from this week's tickets: [paste], and last week's counts: [paste], write a one-page report for the product team: the top five themes, what changed since last week, two customer quotes that show the impact and one recommendation per theme. Plain language, no jargon.
You get: A one-page report product teams will actually read
Write an incident update
We have an incident: [what is broken, who is affected, since when, what we know, next update time]. Write a status page update, a short in-app banner under 25 words and an email to affected customers. Be honest about what we do not know yet, avoid technical blame and give the time of the next update. Do not guess at causes.
You get: Consistent, honest incident messages for every channel
Escalate a ticket clearly
Summarise this customer conversation for an escalation to [engineering / billing / account manager]: [paste, anonymised]. Include the problem in one sentence, steps to reproduce or the facts of the case, what we already tried, the customer impact and urgency and what we need from the other team. Use short bullet points.
You get: An escalation the next team can act on without rereading the thread
Score a reply against our quality checklist
Here is our quality checklist for support replies: [paste]. Score this agent reply against each item, from 1 to 3, with a one-line reason: [paste, anonymised]. Then rewrite the weakest sentence and give the agent one coaching tip. Be constructive.
You get: Consistent quality review and a coaching note
Train a new agent with practice tickets
Write five realistic practice tickets for a new support agent at [company type] selling [product]. Make them increasingly difficult: a simple how-to question, a billing mistake, a frustrated repeat contact, a request we must refuse and a customer who mentions cancelling. Use fictional names. For each, add what a great reply would include.
You get: A training set with model answers