Skip to content
Search
AppliedPlaybooks

AI Customer Support Reply Prompt from Your Own Help Docs

Difficulty
Beginner
Time
10 minutes per reply at first, less once it's routine

Before you start

  • The customer's message, with personal details removed
  • The one to three help articles or policy pages that cover the question
  • An AI chat assistant your company has approved for support work

This AI customer support reply prompt drafts answers from your own help docs, then checks every claim in the draft against those docs. In support, the risk is a confident answer that promises a refund, a feature or a deadline your company doesn't offer. Giving the AI only approved articles, and checking its draft against them, is a simple form of grounding, the same idea behind retrieval-augmented generation.

Step 1: Remove personal details from the ticket

Use the AI assistant your company has approved for this kind of work, and follow your company's AI policy. If you don't know which tool is approved, ask before you paste anything. Unless your approved tool is cleared for customer data, replace the customer's name, email, phone, address, order number and payment details with placeholders such as [NAME] and [ORDER]. You'll put the real ones back in your help desk, not in the AI tool.

Step 2: Pick the help articles

Find the articles that answer the question in your help center or internal knowledge base, and copy their text. If no article covers it, stop here: the answer needs a person, and the ticket-to-article playbook shows how to fill that gap later.

Step 3: Draft the reply

You are drafting a reply for a customer support agent to review.
Use ONLY the help articles below. If they do not answer part of the
question, say "I'll check with the team and get back to you on [that
part]" instead of guessing. Never promise refunds, credits, dates or
features that the articles do not state.

Tone: friendly, plain, short paragraphs. Answer the question in the
first two sentences. Keep placeholders like [NAME] as they are.

Customer message:
[paste]

Help articles:
[paste]

Step 4: Check every claim

Start a new chat so the checker doesn't just agree with itself, and paste this with the draft and the same articles. Using a model to grade text against a source is called LLM as judge; it helps you look in the right places, but it is not the final word.

Below is a draft support reply and the help articles it must be based on.
List every factual statement in the draft, one per line. Mark each:
SUPPORTED (quote the article line), UNSUPPORTED (no article says this),
or CONTRADICTED (an article says otherwise, quote it).

Draft:
[paste]

Help articles:
[paste]

Step 5: Fix, then send it yourself

Remove or rewrite every UNSUPPORTED or CONTRADICTED line. Then read the SUPPORTED lines against the quoted article yourself, because the checker can be wrong too. Put the real customer details back in your help desk, read it once as the customer, and send.

What to do next

Keep a short list of questions where no article existed or the draft kept getting it wrong. That list is the input for turning solved tickets into help articles.

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.

Go further on kief.studio