Using ChatGPT to answer customer emails safely

ChatGPT is good at the wording of a customer reply and bad at the facts in it: it hasn't seen your order, your stock or your returns policy, so when a reply needs one of those it will write something plausible. Use it to word the reply, give it your policy text, leave every order fact as a blank, and read the result before you send it.

Below is what it does well, the three risks of copy-pasting customer emails into it, what OpenAI's own pages say happens to what you paste, and a workflow with a prompt you can copy.

What ChatGPT is good at in a support inbox

Task	Good fit?	Why
Making a blunt reply kinder, or a long one shorter	Yes	It's a writing task; you can see whether it worked.
Replying in a language you don't write well	Yes, with care	Keep it short, and ask it to translate the customer's email back so you can check what they asked.
Turning your returns policy into a plain answer	Yes, if you paste the policy	It can only quote what it has been given.
Saying where an order is, or when it shipped	No	It has no access to your orders. Anything it writes here is a guess.
Deciding whether a customer gets a refund	No	That's your policy and your decision, not a writing task.

The rule behind the table: let it change how something is said, never what is true.

Risk 1: it writes facts it doesn't have

OpenAI says this plainly in its own help centre: ChatGPT "can produce incorrect or misleading outputs. Sometimes, it might sound confident—even when it's wrong." Its list of examples includes "incorrect definitions, dates, or facts". Its advice: "Use ChatGPT as a first draft, not a final source."

In a support reply the facts most likely to be made up are the ones the customer cares about most: a delivery date, a tracking number, the length of your returns window, whether shipping is refunded. Asked "when will my parcel arrive?", it doesn't answer "I can't know that"; it writes a date. For real cases where support AI invented a policy, see AI hallucinations in customer service.

Risk 2: customer data leaves your inbox

A customer email contains a name, an email address, often a postal address and an order number. When you paste it into ChatGPT, that text is sent to OpenAI. What happens next depends on your plan. Here is what OpenAI's pages say, quoted:

Plan	What OpenAI says about training
Free, Plus, Pro	"When you use our services for individuals (Free, Plus, Pro), we may use your content to train our models." You can turn off Improve the model for everyone under Settings > Data controls.
Temporary chat	A chat "that won't appear in your history, create memories, or help improve our models."
ChatGPT Business, Enterprise	"By default, OpenAI does not use content from ChatGPT Business, Enterprise, Edu, or ChatGPT for Healthcare workspaces to train its models."

Two more lines matter if you sell to customers in the EU or UK. OpenAI says it is "able to execute a Data Processing Addendum (DPA) with customers for their use of ChatGPT Business, ChatGPT Enterprise, and the API"; personal plans aren't on that list. And for ChatGPT Business, "deleted or unsaved conversations are removed from our systems within 30 days, unless longer retention is required by law, or is reasonably necessary to protect our services or any third party from harm".

Turning training off doesn't mean the text stays with you; it still travels to OpenAI to be answered. The simplest protection costs nothing: the wording of a reply doesn't depend on who the customer is, so strip the name, email address, postal address and order number before you paste.

One more thing a pasted email can carry: instructions written for the AI rather than for you. In a chat window that mostly produces an odd answer you can see. In a tool that can act on its own, it's more serious; see prompt injection in customer email.

Risk 3: it promises things you don't offer

ChatGPT wants to be helpful, and in a support reply "helpful" easily becomes "yes": a free return you don't offer, a refund on a final-sale item, an apology voucher. Once that reply leaves your address, it's your promise.

A tribunal has already said so. In Moffatt v. Air Canada, the airline's chatbot told a customer he could claim a bereavement discount after travelling, which its policy page said he couldn't. The tribunal held: "It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot." The case, and what it means for a store, is in our guide Is it safe to let AI answer customer emails?

A safer workflow, step by step

1. Read the email yourself first. Decide what the answer is: refund or not, reship or not, what the order shows. ChatGPT doesn't decide; it words.
2. Open the order in your store admin. Note the facts the reply needs: order number, what shipped and when, tracking, amounts.
3. Strip the personal details. Replace the name, email, address and order number with blanks before you paste.
4. Paste your policy, not your memory of it. Copy the relevant part of your returns or shipping page. If your policy page doesn't answer the question, fix the page; see writing store policies support can quote.
5. Use a prompt that forbids invented facts (below).
6. Fill the blanks from the order, in your own inbox.
7. Read the whole reply before you send it, using the checklist further down.

A prompt you can copy

Write a reply to the customer email below for a small online store. Keep it short, friendly and plain, in British English.

Our decision: {{what_you_decided}}

Use only the policy text below for any rule, time limit or cost. If the policy doesn't answer something the customer asked, write [CHECK] instead of guessing.

Don't write any order number, tracking number, date, amount or name. Leave them as {{order_number}}, {{tracking_number}}, {{ship_date}}, {{amount}} and {{first_name}} so I can fill them in.

Don't offer anything that isn't in our decision or the policy: no discounts, vouchers, free shipping or refunds.

Policy: {{pasted_policy_text}}

Customer email: {{customer_email_without_personal_details}}

Writing the decision yourself is the part that matters most. It turns ChatGPT from the person answering into the person typing.

Check before you press Send

Check	What to look for
Every number and date	Does it match the order in your admin? If it didn't come from the order, delete it.
Every promise	Refund, replacement, discount, free return: is it in your policy, or did you decide it?
Time limits	Is the returns window in the reply the one on your policy page today?
Links	Does each link go to your own site or the carrier's? Remove any you didn't add.
[CHECK] markers	None left in the reply.
Tone	Would you sign it?

If a check fails, fix the reply by hand rather than asking ChatGPT to try again; the second attempt can be wrong in a new place.

Where Corresso fits

Corresso does this workflow inside a shared support inbox. Every incoming email gets a reply drafted from documents your store approved (returns policy, shipping page, FAQ), and each fact in the draft cites the document it came from. Order facts are blanks: with Shopify connected, the customer's orders appear beside the email and one click fills the order number, tracking details and ship date. The AI never sees order data. Rules written in code hold a draft that promises a refund no approved document allows, or contains an order detail the AI wrote itself, and they run again after your edits. A person always reads the draft and presses Send. How AI drafts work in Corresso.