Playbook
Customer replies from a written FAQ, drafted in ChatGPT
Direct answer
Put your real refund, price, and booking rules into a ChatGPT Project and let it draft customer replies from that page only. You send the email. If the answer is not in the page, the draft should say you will check, not invent a policy. Not live-tested yet.
- Time to set up
- 1–2 hours, mostly writing the policy
- Difficulty
- beginner
- Monthly cost, USD
- $0–$20
- Monthly cost, GBP
- about £0–£15
- Last tested
- Not yet tested
- Updated
- 24 October 2026
ChatGPT's free plan may be enough for a handful of drafts a day. Plus was listed at $20 a month before tax in October 2026, about £15. The policy page itself costs nothing. Confirm the pricing page, because plan names move.
GBP figures are a budgeting conversion at £1 = $1.33 and €1 = £0.86, rounded to the nearest pound, fixed for this edition on 24 October 2026. They are not a live rate and not a quote. Vendor list prices are usually quoted before tax.
Primary tool: ChatGPT · Also uses Gmail
Who it is for
A small business that answers the same questions every week and can write those answers down. Not a substitute for a complaints handler.
The hour you spend writing the policy is the setup. The model is the photocopy. If you cannot face writing the refund rule, you are not ready for a reply agent, because the agent will write one for you and a customer will rely on it.
Read ChatGPT against Claude for this job if you are choosing the product rather than the habit.
What you need
- A written page of prices, booking rules, and refund rules you are willing to stand behind.
- A mailbox you send from yourself.
- A ChatGPT Project you can wipe when the rules change.
Step by step
01
Write the page customers should have read
One document: what you sell, the prices you will publish, where you work, how booking works, how cancellations and refunds work, how long a reply usually takes, and what you will not do. If you do not have a refund rule, decide one before you ask a model to explain it. A blank is better than a rule you invent under pressure.
02
Put only that document in a Project
Name a ChatGPT Project Customer replies. Instructions: answer only from the attached page; if it is not covered, say you will check and come back; never offer a discount, a compensation figure, or a legal conclusion; use short sentences; add a line that starts with Not in the page. When the policy changes, replace the document and delete old chats so last year's price cannot leak back in.
03
Draft from the real message
Paste the customer's email. Read the Not in the page line first. If the gap matters, answer it yourself and then add a sentence to the policy so the next draft is covered. Send from your own mailbox. Do not connect the Project so that it replies on its own, even if a new screen offers an inbox action.
04
Keep a do-not-draft list
Messages that mention injury, a solicitor, a chargeback, discrimination, or a safety problem get no model draft. You handle them, or you ask someone qualified. Put that instruction at the top of the Project as well. On a busy day the list in your head is the first thing you skip.
05
Review five sent replies a week
Compare what you sent with the policy. If you correct the same sentence repeatedly, fix the policy. If customers still write because the website is unclear, change the website. A faster wrong answer is not an improvement. Stop using the Project for a topic it keeps missing.
What can go wrong
A refund rule that was never yours
Consumer-facing businesses in the UK already have legal duties. A model summary of those duties, written into a customer email, can overpromise or underplay them. Your page should state your process. A lawyer or a trusted adviser is who interprets the statute, not the Project.
Old prices left in the chat
Projects remember what you leave in them. After a price change, search the instructions for the old number. One leftover sentence is enough to quote it.
Auto-reply while you are closed
An unattended reply over a weekend can confirm a booking or a refund you did not mean. This playbook is a draft queue, not an out-of-office bot.
UK notes
Costs, including VAT
About £0 to £15 a month at the planning rate, before any VAT on the ChatGPT invoice. There is no extra software required. If a checkout adds UK VAT, the non-registered business bears it. A VAT-registered business buying services from overseas should check whether the reverse charge applies and keep the invoice.
Data flow
The customer's email often contains a name, an order, and a complaint. Pasting it into ChatGPT sends personal data to OpenAI. Say so in your privacy notice if this becomes a regular process. Do not paste card numbers, identity documents, or anything about health or children. OpenAI's privacy policy is the vendor's description of what it does. You still choose what to send.
HMRC and Making Tax Digital
Nothing here files a return. Refunds you actually pay may affect your records, and those records belong in your bookkeeping, not in the chat. Making Tax Digital does not require an AI reply tool. If you are in it, the reply agent is irrelevant to the quarterly update except as a possible software expense.
Practical orientation for a UK reader. Not tax, legal, or accounting advice.
Test log
Status: Pending first test
24 October 2026 · pending
Pending first test. The steps were compiled from how ChatGPT Projects are documented and from ordinary mailbox practice. They have not been run against a live customer inbox.
Sources
- ChatGPT pricing — accessed 24 October 2026
- OpenAI privacy policy — accessed 24 October 2026
- ICO guide to UK GDPR — accessed 24 October 2026
Disclosure
AgentShed is independent and this playbook is not sponsored. There is no affiliate link on this page. If that changes, the link will use rel="sponsored" and a sentence beside it will say so. The verdict is not for sale.
Questions people ask
Is a website chatbot better?
Often no. A clear FAQ on the site prevents the email. A chatbot that can invent policy is a new way to make the same mistake in public. Drafts in your mailbox are easier to stop.
Claude or ChatGPT?
See the comparison. Claude tends to be the better fit for a long policy. ChatGPT is the better fit if you already pay for it and the rules are short. Neither should send.
What if the customer is wrong?
The draft can point at the written rule. It should not argue, diagnose their motives, or offer a gesture you have not authorised. You can still decide to make an exception. That decision is not in the Project.
Was this useful?
Related playbooks
Compare the tools
When this page is re-tested
One email when the log changes. No sequence of lessons.