Which ChatGPT prompts work for customer service?
The prompts that work give ChatGPT three things: your policy, the customer's exact words and a limit on what it may promise. Without the policy, it invents a refund window. Without the limit, it apologizes for 200 words. The 20 below are the ones I'd hand a new agent, grouped by ticket type, and each one ends with the rule that stops its usual failure. Helpdesk AI agents that answer on their own bill $0.49 to $2.00 per ticket on their October 2026 price pages; a prompt an agent pastes costs nothing per ticket.
| Ticket type | Paste this, then the customer's message | The rule that stops the usual failure |
|---|---|---|
| 1. First reply to a new ticket | "Reply to this customer as a support agent for [company]. Answer every question they asked, in the order they asked it, and confirm the next step with a date. Under 120 words." | "Answer every question" stops the reply that skips the second one |
| 2. Refund approved | "Write a reply confirming a refund of [amount] to [payment method]. Say when it will arrive ([5 to 10 business days]) and what they need to do: nothing, or [step]. Under 80 words." | Give the timeline yourself; the model guesses one otherwise |
| 3. Refund refused | "The customer asked for a refund outside our policy: [paste policy]. Write a reply that says no in the first two sentences, quotes the policy line, and offers [alternative]. No apology for the policy itself." | "Say no in the first two sentences" stops the buried refusal |
| 4. Late or lost order | "Order [number] shows [carrier status]. Write a reply that states what we know, what we don't, the date we'll update them by, and what happens if it hasn't arrived by [date]." | "What we don't know" stops a confident false delivery date |
| 5. Angry customer | "This customer is angry. Write a reply that names the specific problem in their words, takes responsibility for our part only, gives one concrete fix and one direct contact. No phrases like we value your feedback. Under 100 words." | Ban the stock phrases by name |
| 6. Customer threatens to cancel | "The customer says they will cancel because [reason]. Write a reply that fixes [reason] if we can, states plainly what we can't change, and tells them how to cancel if they still want to. Do not offer a discount." | "Do not offer a discount" stops the model giving away margin |
| 7. Missing information | "Write a short reply asking for exactly these details: [list]. Explain in one sentence why we need them. Number the questions so they can answer in one message." | Numbered questions get one reply instead of three |
| 8. Our mistake or an outage | "We caused this: [what happened, when, who was affected]. Write an apology that says what went wrong, what we fixed, and what we're doing so it doesn't repeat. No blame on any vendor. Under 120 words." | State the facts first; the model fills gaps with excuses |
| 9. Handoff to tier 2 | "Summarize this thread for a tier 2 agent: customer goal, what was tried, what failed, account details mentioned, and the one question tier 2 must answer. Bullet points, no greeting." | "The one question" saves tier 2 rereading the thread |
| 10. Bug report for engineering | "Turn this ticket into a bug report: steps to reproduce, expected result, actual result, device and version, how many customers reported it. Write not stated for anything missing." | "Not stated" stops invented reproduction steps |
| 11. CRM note | "Summarize this ticket in three lines for our CRM: the issue, the resolution, any follow-up owed and its date." | Three lines, or it writes a paragraph |
| 12. Shorten a reply | "Cut this reply to [60] words. Keep every number, date, name and next step. Remove apologies after the first one." | List what must survive the cut |
| 13. Tone check | "Read my reply before I send it. Flag any sentence that sounds defensive, sarcastic or vague, and suggest one rewrite each. Don't change anything else." | "Don't change anything else" keeps your reply yours |
| 14. Translate a reply | "Translate this reply into [language] for a customer. Keep it as formal as the English. Then translate your version back into English so I can check it." | The back-translation is your check |
| 15. Saved reply (macro) | "Here are five replies our best agents sent for [issue type]. Write one saved reply that does what they all do, with [brackets] for the parts that change per customer." | Real replies beat a model's idea of a good one |
| 16. Help center article | "Turn this solved ticket into a help center article: a title phrased as the customer's question, the answer in the first sentence, then numbered steps. Remove all customer details." | "Remove all customer details" before it goes public |
| 17. Live chat quick answers | "Write [5] live chat replies for [issue], each under 30 words, friendly and plain. One must hand off to a human agent." | The handoff line is the one a bot forgets |
| 18. Callback script | "Write a phone script for calling this customer back: greeting, the one-sentence summary of their issue, the fix, a check that it worked, and a close. Mark where I pause for their answer." | Pauses written in, or it reads like a monologue |
| 19. Public review reply | "Write a public reply to this [2 star] review. Thank them for one specific thing, answer the complaint in one sentence, invite them to contact [channel]. Under 70 words. Never confirm order or account details in public." | Nothing private in a public reply |
| 20. Weekly ticket themes | "Here are [N] ticket subjects and first lines from this week, with names removed. Group them into themes, count each theme, and list the three that grew most against last week's list below." | Counts, not adjectives, so the report can be checked |
The bracketed parts are yours. The prompts work in ChatGPT, Claude or Gemini without changes, because none of them depends on a feature only one app has. If you also write outbound mail, AI for email writing covers the cold and follow-up side.
Does AI actually make customer service agents faster?
Yes, by 15% on average and by far more for new agents, in a field study of 5,172 agents at one support desk. Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed 5,172 agents at one software company as an AI assistant rolled out in stages, and published the result in the Quarterly Journal of Economics in February 2025. The assistant suggested replies that agents could take, edit or ignore, which is what the 20 prompts above do by hand.
The paper's own summary is blunt: "Access to AI assistance increases worker productivity, as measured by issues resolved per hour, by 15% on average." Agents with two months on the job and the assistant performed about as well as agents with more than six months and no assistant. Average handle time fell about 3.7 minutes from a 43 minute baseline, roughly 8.5%. Customers changed too. The authors found "customers are more polite and less likely to ask to speak to a manager."
Two results cut the other way, and they're the ones to plan around. The resolution rate barely moved, up 1.3 points from 82% and not statistically significant. The most experienced and highest skilled agents saw "small gains in speed and small declines in quality." So give the prompts to new hires first. Let your best agents write the saved replies that prompt 15 learns from.
Should a support team buy an AI agent or paste prompts into ChatGPT?
Paste prompts until the same questions arrive faster than your team can answer them, then price an AI agent per ticket. An AI agent answers customers on its own and bills each conversation it handles. A chat subscription drafts for an agent who reads and sends, and it costs the same whether that agent handles 50 tickets or 500. I read each vendor's pricing page on 25 October 2026, US list prices.
| Helpdesk AI agent | Price per AI-handled ticket | What counts as one | Cost of 1,000 a month | Seat price for human agents |
|---|---|---|---|---|
| Freshdesk Freddy AI Agent | $49 per 100 sessions, about $0.49 each | A session; first 500 included once per account | about $490 | Growth $19 an agent a month, billed yearly |
| Gorgias AI Agent | $0.90 on annual plans, $1.00 monthly | An automated interaction, which also counts as a ticket | $900 to $1,000 | Needs a Gorgias helpdesk plan, priced separately |
| Intercom Fin | $0.99 per outcome | The customer confirms it's resolved, stops asking, or Fin completes a workflow | $990 | Essential $29 a seat monthly, $19 yearly |
| Zendesk AI agents | $1.50 committed, $2.00 pay as you go | An automated resolution; 5 to 10 included per agent a month | $1,500 to $2,000 | Suite Team $55 an agent a month yearly, $69 monthly |
| Salesforce Agentforce | $2 per conversation | A conversation; the page doesn't say it must be resolved | $2,000 | Priced separately |
A chat plan sits on the other side of that math. Claude Pro is $20 a month, or $17 billed yearly, and ChatGPT Business is $25 a seat billed monthly or $20 yearly, with a 2-seat minimum. Five agents on chat seats cost about $100 to $125 a month in total. That's a tenth to an eighth of what Intercom Fin bills for 1,000 resolutions, and the agents still read every reply before it goes out. The AI agent wins when the same questions repeat every day and your help center already answers them, because it replies at 3 a.m. on a Sunday with nobody on shift and bills only for what it handles. Most small teams aren't there yet.
This page is published by Whizi, which sells the chat plan side: 280+ models on one login. Starter is $15.99 a month, or $10.99 a month billed yearly, for ChatGPT Luna and Claude Haiku. Gemini and side by side chats start on Pro at $29.99. Whizi loses to every row in that table on one thing. It has no helpdesk integration and no team plan, so it can't answer a customer by itself or share a seat across a support team. It fits the agent who wants to run prompt 3 or prompt 8 past two models before a hard reply goes out.
How do you give ChatGPT your policies so replies stay accurate?
Write one support brief and paste it ahead of the prompt, or save it where the app rereads it on every message. Most wrong replies come from a missing fact, not a bad model. The brief runs 150 to 300 words and covers six things:
- What you sell and to whom, in two sentences.
- Your refund, return and cancellation rules, quoted from the policy page.
- Shipping or service times you can stand behind.
- Your tone in three words and two phrases you never use.
- What agents may offer without a manager (a credit, a replacement, nothing).
- One real reply your team liked, pasted whole.
Then add one line to every prompt: "Use only the facts in the brief. If the customer asks something the brief doesn't cover, write [CHECK] instead of guessing." The [CHECK] tag is the cheapest safety net in support work. It turns a confident wrong answer into a visible gap that the agent fills before sending. ChatGPT and Claude both have project spaces that hold a brief like this, and the prompts get shorter once it's saved.
What should a support agent never paste into ChatGPT?
Never paste card numbers, passwords, government ID numbers or a customer's health details. The prompts above don't need them. Replace the name with [customer] and the order with [order], and the draft comes back just as useful. Check your company's rules too. Many support teams forbid pasting tickets into a personal chat account at all, and a business plan with an admin is how they allow it.
Watch prompt 20 most closely. A week of tickets is a pile of personal data, so export subjects and first lines only, with names and emails stripped first. What never to paste into AI has the full list and the settings that control what each chat app keeps. If your team is choosing between a personal and a business plan, ChatGPT Business vs Plus compares the admin and data terms.
Frequently asked questions
Can ChatGPT be used for customer service?
Yes, as a drafting tool for agents, and that's where the evidence is strongest. In a 5,172-agent study published in 2025, AI suggestions raised issues resolved per hour by 15%. An agent still reads and sends each reply. A bot that answers customers alone is a separate product, priced per ticket by helpdesk vendors.
What is the best prompt for customer service replies?
The best prompt pastes your policy and the customer's exact message, asks the model to answer every question in order, sets a word limit and bans promises the policy doesn't make. Prompt 1 above is that shape. For harder writing work, is Claude Pro worth it compares Claude's paid plan against the free tier.
Will AI replace customer service agents?
Not on the evidence so far. The 5,172-agent field study found AI helped new agents most and changed little for veterans, and the resolution rate barely moved. Vendors now sell AI agents that close simple tickets alone for $0.49 to $2.00 each, so the routine questions shift to bots while people keep the hard ones.
Is it safe to paste customer emails into ChatGPT?
Only with personal details removed, and only if your company allows it. Strip names, emails, card numbers and account numbers, and the prompts work just as well. A business plan with an admin is the usual way teams permit it. What never to paste into AI lists the data to keep out.
