Drafting Customer Responses with AI

This is the lesson that will save you the most time every day. Drafting replies is the bread and butter of customer support, and AI is astonishingly good at it once you know how to prompt. By the end of this lesson you'll draft polished, on-brand, customer-ready replies in about 20% of the time you spend now.

What You'll Learn

• The four-block response drafting prompt that works every time
• How to handle refunds, apologies, bug reports, and cancellations with AI
• How to use AI to write multi-part replies (acknowledge + explain + next step)
• The "never-send" cases where AI drafts shouldn't leave your draft folder

The Four-Block Prompt for Drafting

The fastest way to a great draft is a prompt with four blocks: situation, policy, tone, and output spec.

Situation: A customer named [Name] wrote in saying [what they said]. They're a [plan/tier] customer who signed up [when]. Their order/account is [relevant ID].

Policy: Our return window is [X days]. For defects we ship a replacement + prepaid return label. Refunds to credit cards take 5-7 business days.

Tone: Warm, apologetic but not groveling. Use contractions. Start by acknowledging their feeling. Use their first name once.

Output: An email reply, under 130 words, with a 6-word subject line.

Paste this into Claude or ChatGPT and you'll have a ready-to-edit draft in seconds. The four blocks work because they match how humans actually think: what happened, what are the rules, how should I sound, what am I producing.

Refund Requests

Refund replies are where agents waste the most time trying to sound natural while being firm on policy. A sample prompt:

Situation: A customer wants a full refund on a 40-day-old purchase. Our return window is 30 days. She says her father passed away and she never had time to try the product. Tone is grieving but not angry.

Policy: Return window is 30 days; exceptions for bereavement are allowed on a case-by-case basis at the agent's discretion, up to 60 days.

Tone: Compassionate. Warm. Do not mention the policy explicitly -- just grant the exception.

Output: Email reply, under 120 words. Offer a full refund with prepaid return, express sympathy without being performative.

The resulting draft sounds like a human being, not a policy bot. Your edit is 10 seconds.

Apology Replies

When your company messed up -- shipped the wrong item, missed a deadline, billed incorrectly -- a great apology is three parts:

1. Acknowledge what went wrong, specifically.
2. Fix it -- what you're doing right now to resolve.
3. Repair -- what you're offering as a gesture of goodwill.

Prompt it like this:

Draft an apology email for a customer whose order #45892 shipped 9 days late because of a warehouse error. Use the acknowledge-fix-repair structure:

• Acknowledge: name the specific mistake (warehouse delay)
• Fix: the package arrived yesterday; confirm they have it
• Repair: offer a $15 credit applied to their next order

Keep it under 110 words, warm, no corporate phrases, sign off as [your name].

Bug Report Replies

Bug replies are tricky because you often don't know when it'll be fixed. Use this pattern:

Draft a reply to a customer who reported [specific bug]. Our engineering team has confirmed the bug and is working on it. We don't have a specific ETA.

Structure:

1. Thank them for the detailed report
2. Confirm we've reproduced the issue internally
3. Set honest expectations: we don't have a fix date yet, but we'll email them as soon as one ships
4. Offer a workaround if they ask: [workaround if any]

Tone: appreciative, honest. Never promise a specific date.

The "never promise a specific date" constraint is important -- AI will invent a date if you don't tell it not to.

Cancellation Replies (Retention)

When a customer asks to cancel, you have about 30 seconds to say something that either retains them or leaves them with a good taste. Both are valuable.

Draft a cancellation reply for a customer who has used our product for 8 months and is cancelling because "we've moved to a competitor that has better reporting features."

Structure:

1. Acknowledge the reason without being defensive
2. Mention our new reporting dashboard launching next month -- ask if they'd like to stay and try it
3. If they still want to cancel, confirm you'll process it today and their data will be exported to them

Tone: gracious, not pushy. If they want to leave, let them leave well.

A gracious cancellation reply gets you good word-of-mouth from a customer who might otherwise badmouth you.

Multi-Part Replies

Sometimes customers ask three things in one email. AI is great at structured responses:

A customer asked three things in one email:

1. Why is my invoice $15 higher this month?
2. Can I pause my account for a month?
3. When will the mobile app support dark mode?

Draft a reply that answers each in order, using a numbered list. Answers:

1. Price increased because they upgraded on the 15th
2. Yes -- we offer pauses up to 90 days
3. Dark mode ships in the next app update, no date yet

Tone: warm and professional. Under 160 words total.

Structured replies score higher on CSAT because customers feel heard on every question, not just the first.

Never-Send Cases

Some replies should never be AI-generated without heavy human rewriting:

• Condolences for death, illness, or trauma -- AI language often feels performative
• Legal matters -- lawsuits, GDPR data deletion, regulatory complaints
• Security incidents -- any mention of account compromise, unauthorized access, data breach
• Account-specific numbers -- refund amounts, balances, dates. AI will sometimes invent numbers
• Public-facing replies on Twitter/Trustpilot -- higher scrutiny, single wrong phrase goes viral

For these, use AI to draft a skeleton and then rewrite heavily in your own voice.

Workflow: Edit to 20% of Drafting Time

Here's the workflow that works:

1. Open ChatGPT or Claude in one tab, your help desk in the other.
2. Copy the ticket into your prompt template.
3. Read the output. Scan for invented facts or off-tone sentences.
4. Edit 1-2 sentences to make it feel like you.
5. Verify any numbers, URLs, or policies against source truth.
6. Send.

Agents using this flow report drafting time dropping from ~4 minutes per ticket to ~50 seconds.

Key Takeaways

• Use the four-block prompt: Situation, Policy, Tone, Output
• Apologies follow acknowledge-fix-repair structure
• Never let AI invent dates, refund amounts, or policy numbers -- paste them in
• Gracious cancellation replies are retention opportunities
• Legal, medical, and condolence replies need heavy human rewriting before sending