ChatGPT for customer service

How do you use ChatGPT for customer service?

Treat it as a very fast writer sitting next to whoever answers customers. It doesn’t plug into your inbox or your order system by default, and it doesn’t reply to anyone on its own. You paste a message in, it drafts the answer, a person checks it and sends it.

Sounds modest. In practice that’s most of the writing in a support job, and the part people dread. This page is part of my AI assistants series, and if you’re still choosing a plan for the wider business, start with ChatGPT for small business.

The one rule: ChatGPT doesn’t know your refund policy. Left alone, it’ll happily invent a generous one. Everything below is about stopping that.

How to set up ChatGPT for customer service

Skip the one-off chats. The setup that works is a Project. OpenAI’s Projects docs say each project has a Sources section for uploaded files and connected context, and that project instructions apply across its chats. So you load your policies once, and every reply drafted inside the project follows them.

Paste this into the project instructions and fill in the brackets:

You draft customer service replies for [business name], which sells [what you sell] to customers in [country/region].

Facts:
- Use only the policies, prices and dates in this project's sources. Never invent a policy, discount, refund amount or delivery date.
- If the answer isn't in the sources, write [CHECK: what you need to confirm] instead of guessing.

Tone:
- Plain US English, friendly, short. Max 120 words unless asked.
- Use the customer's first name. No "We apologize for any inconvenience."
- One clear next step at the end.

Never:
- Promise a refund, replacement or compensation unless I say so in the message.
- Blame the customer, the courier or a supplier by name.
- Ask for card numbers or passwords.

The [CHECK] line is the one I’d never remove. It turns a confident guess into a visible gap you fill in before sending.

Copy-ready ChatGPT prompts for customer service

Run these inside your Customer service Project, so ChatGPT already knows your policies. Swap anything in square brackets. Notice that every prompt hands it the decision. ChatGPT writes the reply, you decide the refund.

1. Refund request

Customer message: "[paste]"
Order facts: ordered [date], delivered [date or not yet], item [name], paid [amount].
Our decision: [full refund / partial refund of $X / store credit / no refund because ...].

Write a reply that states the decision in the first two sentences, explains it in one sentence using our refund policy from the sources, and says exactly what happens next and when.
If the decision is no, offer the one alternative our policy allows.

2. Late delivery

Customer message: "[paste]"
Tracking status: [paste status and last scan date]. Normal delivery time: [X business days].

Write a reply that says where the parcel is in plain words, gives a realistic date range (not a promise), and tells them what we'll do if it hasn't arrived by [date].
Don't blame the courier. Keep it under 100 words.

3. Angry customer

Customer message: "[paste]"
What actually happened: [one or two lines of facts].
What I can offer: [options, or "nothing beyond the policy"].

First, list in one line each: what they're upset about, what they want, and anything in their message that's factually wrong.
Then write a reply that acknowledges the frustration in one sentence, corrects any wrong facts politely, and gives a concrete next step.
No groveling, no corporate phrases, under 130 words.

4. FAQ macro builder

Below are our 20 most common customer questions.
For each one, write a saved reply (macro) for our help desk:
- A short title for the macro.
- A reply under 80 words, using only facts from this project's sources.
- Placeholders in {{double braces}} for anything that changes per customer, like {{first_name}} or {{order_number}}.
Mark any question you can't answer from the sources with [CHECK].

Questions:
[paste list]

Public replies need extra care, because everyone sees them. My guide on how to respond to Google reviews covers tone and what never to say.

The angry-customer prompt is the one I’d start with. Splitting “what they want” from “what they said” first gives a much calmer draft than asking for a reply straight away.

Is customer data safe in ChatGPT?

It depends on the plan. Sort this out before anyone pastes a real customer email, not after. On chatgpt.com/pricing, the individual plans (Free, Go, Plus, Pro) say content is used to train OpenAI’s models, with an opt-out available. The Business plan on OpenAI’s pricing docs lists “No training on your business data by default,” plus a dedicated workspace with admin controls, SAML SSO and MFA. Checked Nov 2026.

My rule: one person on Plus with training switched off in Settings > Data controls is fine for a start. Once two or more people handle customer messages, move to Business.

Whatever the plan, strip what the reply doesn’t need. ChatGPT doesn’t need the customer’s card number to apologize for a late parcel.

When a help-desk AI agent beats ChatGPT

The ChatGPT app makes your team faster. It doesn’t answer customers while you sleep. For most small companies that’s fine. For that you need an AI agent inside a help desk, which reads your help center, replies on chat or email, and hands off to a person when it’s stuck. These are the main options, from each vendor’s own pricing page, checked Nov 2026:

Intercom’s Fin pricing is the cleanest example of the pay-per-answer model. Its FAQ says resolutions, procedure handoffs, disqualifications and self-serve routings are $0.99 each, charged once per conversation.

Zendesk’s pricing page says its AI agents are “included in every Suite and Support plan, with pricing based on the successful outcomes they deliver.” Copilot, its assistant for human agents, is an add-on at $50 per agent a month paid yearly. Freshdesk offers its Freddy AI Copilot at $29 per agent a month on Pro and Enterprise. Gorgias says its help desk “scales from 50 to 5,000 tickets a month, never priced per agent.”

Here’s what a two-person support team pays a month on each route. The Intercom bar assumes 200 Fin outcomes a month at the lowest seat price; Zendesk and Freshdesk AI usage beyond their allowances costs extra.

The math is simple. Fin costs more than a ChatGPT seat once it handles a few dozen conversations a month, but every one of those is a conversation nobody on your team touched. If your people spend hours a day on “where’s my order,” the agent wins. If you get 15 emails a day, ChatGPT and a good set of macros is plenty, and buying an agent is solving a problem you don’t have yet.

An AI agent that answers at 2 a.m. is great. One that promises a refund you don’t offer is an expensive pen pal. Whichever you pick, feed it the same policy pile from step 1.

Other ways to wire ChatGPT into support

A few routes sit between “paste into ChatGPT” and “buy an AI agent”:

• Connect your tools. If your help desk or store has an MCP server, you can add it to ChatGPT as a plugin so it can look things up, see ChatGPT MCP.
• Automate the drafting. A Zap can send each new ticket to an AI step and drop the draft back in as an internal note, see Zapier AI automation.
• Build your own agent. OpenAI’s Agent Builder was the obvious DIY route, but it’s shutting down on November 30, 2026, see OpenAI Agent Builder before you start anything there.
• Phone calls. ChatGPT doesn’t pick up your phone. For that, compare AI receptionist pricing.
• A different assistant. Claude handles long policy documents and careful tone well. The setup is the same, see Claude for small business.

If you’d rather someone set the whole thing up for you, some AI automation companies do help-desk AI projects. I’d start with the Project and the prompts above anyway. A month of that and you’ll know exactly what to ask for, and what not to pay for.