AI for Customer Service Replies: Draft From Your Own Help Docs
TL;DR: Paste the help article that covers the question, ask the AI to answer only from that text, and read the draft before you send it. You'll reply faster, and the answer will match what your company actually offers.
Using AI for customer service replies is one of the most practical ways to start with AI at work. Much of support work is writing: reading what a customer asked, finding the policy or help article that answers it, and turning that into a clear, kind reply. An AI assistant is good at the last step. It's much less reliable at the middle step, finding the right facts, unless you hand those facts to it.
This lesson shows you how to do that safely. You don't need to install anything. You need the AI assistant your company approves, your own help docs, and a few minutes to try the prompt below on a real ticket.
Why this works for support teams
There's solid research behind using AI in support. A study of 5,172 customer support agents at one company found that access to an AI assistant raised the number of issues resolved per hour by 15% on average (Brynjolfsson, Li and Raymond, "Generative AI at Work", The Quarterly Journal of Economics, 2025). The gains weren't even. Less experienced agents got faster and better, while the most experienced agents saw small gains in speed and small declines in quality.
That second finding is the useful one for you. The AI helped most where people didn't yet know the answer by heart. It also means an experienced agent shouldn't accept a draft just because it reads well. Your judgment is still the quality check.
The one problem you have to plan for
AI models can write a confident answer that is simply wrong, such as a refund window your company never offered or a feature your product doesn't have. This is called hallucination, and in support it matters a lot, because the customer may act on what the reply promises.
In support, a hallucination usually looks like a promise. The customer reads it as your company's word, because it came from your company.
The fix is a habit called grounding. Grounding means you give the model the source text and tell it to answer only from that text. If the answer isn't in the text, the model should say so instead of filling the gap. When software does this automatically by searching your help center first, it's called retrieval-augmented generation, or RAG. You can do the manual version today by copying and pasting.
Before you paste anything: what stays out
Follow your company's AI policy first. If your company hasn't approved an AI tool, ask before you use one for customer work. Then strip the ticket before it goes in.
Never paste:
- Customer names, email addresses, phone numbers or home addresses
- Order numbers or account IDs that link back to a person
- Card numbers, bank details or anything from a payment screen
- Passwords, login links, API keys or internal system screenshots
- Health, legal or other sensitive details the customer shared
You don't need any of that to draft a reply. Replace it with placeholders such as [CUSTOMER NAME] and [ORDER NUMBER], and fill them back in when you send. NIST lists data privacy, meaning leakage or unauthorized disclosure of personal information, as one of the core risks of generative AI (NIST AI 600-1). Placeholders are the simplest way to stay clear of it.
The prompt: answer from the help doc only
Here's the prompt to try on your next ticket. Paste the customer's question (with personal details removed) and the help article or policy text that covers it.
The Prompt:
You are helping a customer support agent draft a reply.
Rules:
1. Use only the facts in the HELP TEXT below. Do not add policies, dates, prices or features that are not in it.
2. If the HELP TEXT does not answer part of the question, write [NEEDS AGENT] at that point and say what is missing.
3. Keep placeholders such as [CUSTOMER NAME] exactly as written.
4. Write in a friendly, plain tone. Short paragraphs. No more than 150 words.
5. After the reply, list each fact you used and quote the sentence from the HELP TEXT it came from.
CUSTOMER QUESTION:
"""
Hi, I bought the large plan two weeks ago and want to switch to the small one. Do I get money back for the difference? Order [ORDER NUMBER].
"""
HELP TEXT:
"""
[paste the help article or policy section here]
"""
Why This Works: Rule 1 limits the model to your own words, which is what grounding means. Rule 2 gives it a safe way to admit a gap, so it doesn't invent an answer to look complete. Rule 5 makes it show its sources, so you can check each claim in seconds instead of rereading the whole help article.
What you get back: A short reply in your tone, with a [NEEDS AGENT] marker anywhere the help text is silent, followed by a list such as "Plan changes take effect at the next billing date: quoted from paragraph 2." You check the list, fill the gap yourself or ask a colleague, add the customer's details back in and send.
Check before you send
Read every draft as if you'll be held to it, because you will. Three checks help you catch problems:
- Match each promise to a quote. If the reply promises money, time or a feature, the source list should show the sentence it came from. No quote means the claim isn't backed by your help text, so delete it.
- Look for [NEEDS AGENT]. Those markers are the parts you still have to answer. Don't send a reply with one left in.
- Read it out loud once. If it sounds stiff or overly apologetic, say so in a follow-up message to the AI ("shorter and warmer, keep the facts") or fix it by hand.
If the same question keeps hitting [NEEDS AGENT], the help article itself has a gap. Tell whoever owns your help center. That's one of the quiet benefits of drafting this way: it shows you where your documentation is thin.
Three more support prompts to try
Once the first prompt works for you, these follow the same pattern. Each one keeps personal details out and keeps you as the final reader.
- Summarize a long thread for a handover. "Summarize this ticket thread in five bullet points: what the customer wants, what we've tried, what we promised, what's still open and the next step. Use only what's in the thread."
- Turn an angry message into a calm reply. "Draft a reply that acknowledges the problem in one sentence, explains the next step from the HELP TEXT and gives a time frame only if the HELP TEXT states one."
- Suggest a help article update. "Here are five customer questions the HELP TEXT didn't answer. Draft a short new section that would answer them, and mark every sentence I need to confirm with [CHECK]."
The last one turns your daily tickets into better documentation, which then makes every future draft more accurate.
When to stop and hand it to a person
Some tickets shouldn't be drafted with AI at all, or should only be drafted after a lead has looked. Skip the AI draft, or escalate first, when the ticket involves:
- A legal threat, a complaint to a regulator or a safety issue
- A refund, credit or exception outside written policy
- Someone in distress, or a message about a death, illness or emergency
- A security problem, such as a customer reporting a hacked account
In those cases, speed matters less than getting it right, and your company probably has a set process. Follow it.
Where to go next
If you want to see how this works when it's automated, read grounding prompts with retrieved context, which covers the same idea at the scale of a whole help center. For more lessons written for your job, go back to the customer support track, or browse the Applied pillar for AI by role.
Frequently asked questions
Can AI write customer service replies?
Yes, as a first draft. AI is good at turning facts into a clear, polite reply. It isn't reliable at knowing your policies, so give it the help article or policy text and ask it to answer only from that. A person should read every reply before it goes to a customer.
Is it safe to put customer emails into ChatGPT or another AI tool?
Only in the tool your company has approved, and only after you've removed personal details. Replace names, email addresses, order numbers and payment details with placeholders. Your company's AI policy decides which tools are allowed and what data can go in them.
How do I stop AI from making things up in support replies?
Paste the source text and tell the model to use only that text, to mark any gap instead of guessing, and to quote the sentence each fact came from. Then check each promise in the reply against its quote. If a claim has no quote, remove it.
Will AI replace customer service agents?
The study cited here doesn't answer that, because it measured agents working with AI, not AI working on its own. In a study of 5,172 support agents, AI help raised issues resolved per hour by 15% on average. Less experienced agents gained the most, while the most experienced saw small speed gains and a slight drop in quality (Brynjolfsson, Li and Raymond, 2025). That's why this lesson keeps a person reading every reply before it goes out: the AI drafts, and you decide what the customer is told.
What should I never use AI for in customer support?
Don't let it decide refunds or exceptions, answer legal or safety complaints, or respond to someone in a crisis without a person in charge. Never paste passwords, card numbers or account access details into it, even to "check" something.
Rather have someone set it up for you?
That's a fair call. LTFI, the agentic business operating system from Kief Studio, can set up the work these guides describe and run it with you. No website yet? Kief Studio will build you one for free.

