The AI customer service prompt we tested, line by line.
Most customer service prompts you find online are one vague sentence. Ours is one page, and every line is there because of something that happened in a real test: a scam, a refund, a promise the business could not keep. Here it is, with blanks for your business.
Short answer: a good AI customer service prompt tells the AI who you are, which kinds of message you get, how fast each one matters, how you sound, and a short list of things it must never do. Ours is about 220 words. In our test it sorted 10 of 10 emails in 39 seconds, caught a fake $8,750 bank request and wrote every reply as a draft for a human to approve.
The template
Copy it, replace the parts in square brackets, and paste it into your AI assistant before the customer messages. It is the same page we used in episode 1, with our demo business taken out.
You are the email assistant of [Your business], a small [what you do].
For every email I give you:
1. Sort it into one of these: NEW BOOKING, COMPLAINT, REFUND,
CANCELLATION, QUESTION, THANK YOU, SPAM, SUSPICIOUS.
2. Give it a priority: URGENT, TODAY, or LOW.
3. Write a reply draft in our voice: warm, short, human,
no corporate words. Sign as "[The team name]".
4. Reply in the same language the customer wrote in.
Our rules:
- Never send anything. You only write drafts.
A human approves every reply.
- Refund policy: [your policy, in one line].
Never promise more than the policy. Offer a call instead.
- Never give legal, visa, tax or medical advice.
Point to the official source and offer to help.
- Anything asking to change bank details or send money
is SUSPICIOUS: do not reply, flag it for the owner.
- Spam gets no reply.
- Only promise what we actually sell: [your list].
For anything else, say we will check and come back.
- Never say an action is already happening.
Say we have received the request and will confirm.
- If a reply needs information we do not have
(dates, budget, booking number), ask for it.
Output a table with: ID, sender, category, priority,
and one line on why. Then the reply drafts.Prefer a file you can fill in? The free kit has this page plus the ten test emails, so you can see it work before you use it on your own inbox.
Why every line is there
| Line in the prompt | What it did in the test |
|---|---|
| Sort into fixed boxes | 10 of 10 emails landed in a box, with one line on why. You read the table in seconds and know where to look first. |
| Priority: URGENT, TODAY, LOW | A cancellation for the next morning went to the top, a thank you went to the bottom. |
| Reply in the customer's language | An email in Italian got its draft in Italian. |
| Never promise more than the policy | The angry refund email got a calm draft with no money promised and a note to the owner. |
| Bank detail changes are SUSPICIOUS | A fake supplier asked to change bank details and pay $8,750 today. No reply, flagged for the owner. The full story. |
| Only promise what we actually sell | Added after the first run, when the AI offered a party boat the business does not sell. What happened. |
| Never say an action is already happening | Added after the first run, when a draft said "we're arranging the cancellation now" and nobody was. |
| Never send anything | 0 emails sent by the AI. Every draft waited for a person. |
How to adapt it to your business in 15 minutes
- Name and voice. Your business name, what you sell, how you sign off. Three words for your tone are enough.
- Your boxes. Look at your last 30 messages and write down the five to eight kinds you see most. Keep SPAM and SUSPICIOUS: every business gets both.
- Your policy lines. Refunds, cancellations, delivery times: one line each, with real numbers. The AI will quote them, so they must be right.
- Your sell list. Write exactly what you sell. This one line stops most made-up promises.
- Your never list. Start with ours and add your own: never quote a price that is not on the list, never share another customer's details.
Test it before you trust it
Take ten real messages you already answered, run them through the prompt and compare the drafts with what you actually wrote. Look for three things: promises you would not make, actions it says are already done, and anything it answered that it should have flagged. Each problem you find becomes one more plain-English line. That is the whole method: in our test two lines fixed both mistakes, and the second run took 41 seconds.
Common mistakes with customer service prompts
- Only describing the tone. "Be friendly and helpful" makes an AI that says yes to everything. Helpful needs limits.
- No list of what you sell. Without it, the AI fills gaps with whatever sounds nice.
- Letting it send. Drafts first. Keep the send button until you have watched it work for weeks.
- Writing it once. The prompt is a living page. Every mistake you catch is a line you add.
Questions
Does this prompt work in ChatGPT or Gemini?
It is plain English, so any capable chat assistant can follow it. We tested it with Claude, so those are the results we can vouch for.
How long should an AI customer service prompt be?
Long enough to cover your boxes, your policy and your never list. Ours is about 220 words and fits on one screen.
Can I use the same prompt for live chat?
The rules carry over, but our test only covered email drafts approved by a person. A live chat answers customers directly, so test it much more before you switch it on.
Is the template free?
Yes. Copy it from this page or get the kit with the test emails at leeluaistaff.com/kit/ep01/.