Replying to customers well takes longer than it should. The message arrives, you read it, you know roughly what needs to be said, and then you spend twenty minutes finding the right words that are neither too cold nor too casual, and that commit to nothing you cannot deliver. AI handles that drafting step quickly, provided you give it what it needs: the customer's message, the relevant facts or policy, and the tone you want. Without those three things, it guesses, and the guesses get you into trouble.
This guide covers three tasks: drafting a standard reply, handling a complaint, and building a reusable template bank for the queries that come in repeatedly.
The tool, and how to set it up
Claude (claude.ai) works well for this because it follows instruction closely and does not pad replies with phrases customers do not want to read. Paste the customer's message directly into the conversation along with your context. Do not use a persistent Project for live customer replies; keep each conversation clean and isolated.
If your organisation has a Claude for Work or Microsoft 365 Copilot deployment, use those for anything involving customer personal data. Consumer tools are not appropriate for sensitive customer information. More on this in the safety section.
How to do it
For a standard query reply:
- Copy the customer's message exactly as received.
- Write out the relevant facts: what your policy actually says, what you can and cannot offer, what the status of their account or order actually is.
- Decide on the tone before you prompt. "Warm but clear" and "formal, solution-focused" produce different outputs. If you do not specify, Claude picks something middle-of-the-road that may not match your brand.
- Paste the customer message, the facts, and the tone into the prompt. Do not paste just the message and ask it to reply. That produces a generic, fact-free response.
For a complaint:
A complaint reply needs three things: acknowledgement, a substantive response to what went wrong, and a clear next step. Tell Claude to structure it that way. Also tell it what you can and cannot offer. If a refund is available, say so. If it is not, say that too. If you leave it unspecified, there is a good chance it invents a resolution you cannot honour.
Specify the emotional register. A customer who is upset warrants a different tone from someone asking a polite question. "Acknowledge the frustration without being sycophantic" is a useful instruction.
For building a template bank:
Identify the five to ten customer queries that account for most of your reply volume. For each one:
- Draft the ideal response once, using Claude and the method above.
- QA it thoroughly (see below).
- Mark the parts that vary per customer in square brackets: [customer name], [order number], [specific date], [their issue as described].
- Save it somewhere your team can access and fill in quickly.
Review the templates every few months. Policy changes, product changes, and tone drift all make old templates unreliable.
The prompt
You are helping me draft a reply to a customer message. Here is everything you need:
Customer's message:
[paste the full message exactly]
Relevant facts and policy:
[what is actually true: the policy, the status, what we can offer, what we cannot. Be specific. Do not leave this blank.]
Tone:
[e.g. "warm and professional", "direct and solution-focused", "empathetic, the customer is upset", "formal, this is a B2B client"]
Additional context:
[anything else that matters: channel (email, live chat, review platform), relationship with this customer, severity of the issue]
Draft a reply that:
- Addresses what they actually asked or raised
- Stays within the facts and policy I have given you above
- Does not promise, offer, or commit to anything I have not specified
- Matches the tone I described
- Avoids filler phrases ("I hope this message finds you well", "do not hesitate to contact us", "we apologise for any inconvenience")
Do not add anything beyond what I have provided. If there is something you cannot answer from the information I have given, flag it rather than guess.
The last instruction matters. AI fills gaps confidently. Telling it to flag gaps rather than fill them saves you from sending a reply that promises something you never agreed to.
How to QA it
Read the draft twice: once for factual accuracy, once for tone.
Factual accuracy: Check every specific claim against what you actually told it. Does it mention a timeframe you did not provide? A refund amount or policy that does not match yours? A commitment ("we will contact you within 24 hours") you are not able to keep? The classic failure is an invented resolution: AI reads the customer's frustration and helpfully manufactures a solution. That manufactured solution then goes out under your name.
Cross-check the draft against your actual policy document, not your memory of it. Two things that should match exactly: what the customer asked and whether the reply directly answers it, and any commitments in the draft and whether you can honour them.
Tone check: Read it as if you are the customer. If you are upset and you received this, would it land right? Too breezy is a common problem on complaint replies. "Happy to help get this sorted!" is not the right register for someone who has just described a significant problem. If it reads like a chirpy chatbot, ask Claude to make it warmer and more direct, with less service-speak.
Before it goes to the customer: A human reviews and approves it. Not skims it, reviews it. This is not a formality.
How to stay safe
Customer personal data and consumer tools. If the customer's message contains their name, email, order details, payment history, or anything that identifies them, do not paste it into Claude.ai, ChatGPT, or any consumer product without checking whether your organisation's data policies permit it. In most cases, they do not. Use your organisation's approved AI tools, redact the personal details before drafting, or write the reply without AI assistance. A data breach via a poorly thought-through prompt is a real risk, not a theoretical one.
Never send an AI-drafted reply without review. AI replies can sound authoritative while being wrong. They can commit you to policies you do not have, deadlines you cannot meet, or resolutions you never offered. Every draft needs a human check before it reaches the customer.
Template drift. A template that was accurate when you wrote it may no longer be after a policy change. A team using an outdated template confidently is worse than no template at all. Assign someone to own the template library.
Tone mismatches at scale. If you are using AI to draft dozens of replies, it is easy for a house style to emerge that is not quite your voice. Check a sample regularly against how you actually want to sound.
Start with one high-volume query you answer the same way every week. Draft the template, QA it against your policy, and put it somewhere your team can use it. That is the lowest-effort win here.