PRO 35 min read

Lesson 2 — Customer Service That Sounds Like You

Templates and a Custom GPT that drafts customer responses in your voice in 30 seconds. Without making customers feel like they're talking to a bot.

Course Small Business · Lesson 2 of 4

Customer service is the highest-leverage place to use AI in a small business. Done well: customers feel cared for, you reclaim hours every week. Done badly: customers feel like they’re being managed by a robot, and the trust you built falls apart.

This lesson teaches you to do it well.

The mental model

Don’t replace your voice — accelerate it. AI’s job is to draft responses you’ll edit in 15 seconds, not to send replies on its own.

The workflow:

Drag a customer message into AI. AI drafts a response in your voice.
Read it. 15 seconds.
Edit one or two phrases. Make it actually you.
Send.

Total time per response: ~45 seconds. Without AI, the same response (with thought) takes 5–8 minutes.

Step 1 — Build your “response library”

Spend 30 minutes pulling together examples of your best past customer responses. You’re going to feed these to AI as the ground truth for your voice.

Find:

3 responses you wrote when a customer was happy
3 responses you wrote when a customer was upset
3 responses to common questions (refund policy, shipping, product use)

Paste them all into a doc. Label them by tone: HAPPY / UPSET / FAQ.

Step 2 — Build a Custom GPT (or Claude Project)

This is the highest-leverage step. Instead of pasting your voice samples into every prompt, save them once.

In ChatGPT (Plus tier required):

Click “Explore GPTs” → “Create”
Name: “Customer Service Drafter for [Your Business]”
Paste this as the system instructions:

You are the customer service writer for [Your Business]. You draft replies in our exact voice and tone, based on the response samples I'll provide.

Your job:
1. Read the customer message I paste in.
2. Identify the customer's emotional state (happy / frustrated / confused / urgent).
3. Draft a reply that matches our voice and addresses their actual concern.

Rules:
- Match the tone of the response samples exactly.
- Acknowledge before solving — never jump straight to logistics.
- Keep replies short (3–5 sentences) unless the situation needs more.
- Never make promises about timing or refunds I haven't authorized in the samples.
- If the customer needs a manager, flag it: "[ESCALATE: reason]" at the top.
- Never use "Hope this helps!" or "Thank you for reaching out" — they sound corporate.

End every reply with: "[needs review]" so I remember to edit before sending.

VOICE SAMPLES:
[paste your 9 labeled samples]

POLICY NOTES (so you don't promise things we don't do):
- [refund policy]
- [shipping policy]
- [response time expectations]
- [anything else specific to your business]

Save it. Now any time a customer email comes in, paste the message into your Custom GPT — it’ll draft a reply in your voice, ready for a 10-second edit.

(Don’t have ChatGPT Plus? Use Claude Projects — same idea, slightly different setup.)

Step 3 — Master the 5 hardest situations

Most customer messages fit one of these 5 patterns. Sketch your AI’s response approach for each.

Refund / return request

Approach: acknowledge frustration → state policy clearly → offer the best option you can within policy → make it easy to act.

Hi [name], I’m sorry [product] didn’t work out — that’s frustrating. Here’s what I can do: [option A] or [option B]. Just reply with which you’d prefer and I’ll get it set up today.

Angry customer

Approach: don’t be defensive → name what they’re feeling → take responsibility for something (even if it’s just the experience) → propose a concrete fix.

Hi [name], you’re right — [specific thing] shouldn’t have happened that way, and I’m sorry. Here’s what I’m doing: [action]. I’ll [follow-up commitment] by [time]. If that doesn’t work for you, tell me what would.

Confused / “how do I…?”

Approach: answer the question directly first → then briefly explain why → end by inviting follow-up.

Hey [name], short answer: [direct answer]. The reason it works that way is [1 sentence]. If you hit any snag, just reply and I’ll get you sorted.

”I need this urgently”

Approach: acknowledge urgency → set a real expectation → don’t over-promise.

Got it, [name] — I see this is time-sensitive. Here’s what’s possible: I can [realistic action] by [realistic time]. If that won’t work, tell me your hard deadline and I’ll let you know what I can do.

”Can you make an exception?”

Approach: thank them for the context → state the answer with empathy → offer the best alternative.

Hi [name], thanks for explaining the situation — I get it. We can’t [the exception] because [1-line honest reason], but here’s what I can do instead: [alternative]. Let me know if that helps.

Add these 5 patterns as a section in your Custom GPT instructions. Now even hard cases get a strong starting draft.

Step 4 — Set up the daily workflow

Pick one of these:

Option A — Morning batch. Once a day, open all customer emails. Paste each into the Custom GPT. Edit and send each reply in ~45 seconds. 10 emails = 8 minutes total.

Option B — Inline. Each time an email comes in, switch to your Custom GPT, paste, edit, send. Slightly less efficient, but no email backlog.

Use Option A if you have >5 customer emails/day. Use Option B if you have <5.

Step 5 — Quality-check yourself weekly

Once a week, audit 3 random replies you sent. Ask:

Did this sound like me?
Did I solve their actual problem?
Did I leave anything for them to follow up on (and did I follow up)?

If you find 3+ replies in a row that sound flat or AI-ish, your Custom GPT needs better voice samples. Add 2–3 of your best recent replies to the system prompt and re-test.

Common failure modes

“Customers can tell it’s AI.” → Two causes: (1) your voice samples are too few or too generic — add more; (2) you’re sending without editing. Even one human touch per reply makes a big difference.

“It promises things I can’t deliver.” → Add a strict policy section: “NEVER promise [refunds beyond X days / next-day shipping / discounts above Y%].” AI will hallucinate generosity if you don’t pin it down.

“It’s too formal.” → Your voice samples are too formal. Use your most casual reply as a sample and explicitly instruct: “Match the casual tone of sample #2.”

“I don’t trust it for hard cases.” → That’s fine. Use AI for the easy 80% (drafts, FAQs, polite back-and-forth). Handle the hard 20% (escalations, tough refunds, sensitive complaints) yourself. AI is a force multiplier, not a replacement.