How to answer an angry customer without grovelling or getting defensive
The first reply decides whether it calms down or escalates. A two-step prompt that keeps it short, specific and honest.
Written by Tauri on 26 October 2026. About 5 minutes to read.
I wanted a prompt ready before the next angry email lands, not while it's sitting in my inbox. So I tested this one on a practice message: made up, but realistic. A customer charged after cancelling, promised a callback within 24 hours, and still waiting three days later.
The first reply matters most. Get it wrong and you have two problems: the original one and your reply. There are two traps I wanted the AI to stay out of:
- Grovelling. Three apologies, "we deeply regret any inconvenience", and a promise to make it right before anyone knows what went wrong.
- Defensive. A polite paragraph explaining the policy, which to an angry customer reads as "it's your fault".
So I split it into two steps: first work out what they actually need, then write the reply.
Step 1: Separate the anger from the ask
Angry messages mix what happened, how they feel and what they want. Pull them apart before writing anything.
Here's a message from an upset customer: [PASTE MESSAGE, NAMES AND DETAILS REMOVED] Don't write a reply yet. List: - What actually happened, in their words, in one or two lines. - What they want now. If they don't say, write "not stated". - Any facts I need to check before I can answer. Write each one as [CHECK: ...].
Get one tested prompt like this every Tuesday. Sign up below, it’s free.
Step 2: Write a first reply with one job
The first reply doesn't need to solve everything. Its job is to show you've understood the specific problem and to say what happens next.
You help me write my first reply to an upset customer. Its only job is to calm things down and say what happens next. Customer message: [PASTE MESSAGE, NAMES AND DETAILS REMOVED] What I know so far: [FACTS, OR "NOTHING CHECKED YET"] What I can offer right now: [e.g. a call today, a replacement, nothing yet] When they'll hear from me next: [TIME, OR LEAVE EMPTY] Rules: - Start with their specific problem in plain words. Don't open with "I understand your frustration". - Apologise at most once, and only for something I've confirmed went wrong on our side. If I don't know yet, thank them for telling us and say I'm looking into it. - Don't explain policies or defend the company in this reply. - Say exactly what happens next and when. If I haven't given you a time, write [CHECK]. - Only use facts I've given you. If something is missing, write [CHECK]. - Don't promise refunds, discounts or outcomes I haven't listed. - Reply in the customer's language and match their level of formality. Under 120 words. After the reply, list any sentence that could sound defensive or over-apologetic, so I can check it.
The last line is the useful part, the same idea as asking for sources in the note about AI making things up.
What happened when I tried it
Step 1 boiled the practice message down to three lines (charged after cancelling, no callback, wants it sorted today) and three things to check.
For Step 2 I gave it only what I'd realistically know: the cancellation is in the system, I don't know why the charge went through, nobody logged the callback, and I can check the charge today. I left the "when" empty on purpose, to see if it would invent a time.
It didn't. Here's the reply it wrote, with my notes:
Hi [NAME],
You cancelled two weeks ago, you've still been charged, and the callback you were promised on Monday never came.1 That last part is on us, and I'm sorry.2
I can see your cancellation in our system3, dated two weeks ago. I'm checking today why the charge still went through, and I'll call or email you with what I find by [CHECK]4.
- Starts with their actual problem. No "I understand your frustration".
- One apology, for the one thing that definitely went wrong.
- The AI flagged this: it could read like a refund is coming. My call, and I'd keep it.
- The only thing I'd change: a time I can actually keep.
One apology, for the one thing that had definitely gone wrong: the missed callback. No policy explanation, no refund promise. And where it didn't have a time, it wrote [CHECK] instead of making one up.
Its list of risky sentences flagged two things: "I can see your cancellation" could read like a refund is coming, and the reply ignores their threat of a bank dispute. Both are calls for me to make, not things it quietly decided for me.
What I'd change before sending
Honestly, very little. I'd send it. The only thing it needs from me is the [CHECK]: a time I can actually keep.
That's the part I like most. When the AI isn't sure, it tells me what to check instead of guessing, and I end up with a clear view of the next steps.
To be upfront: this was one practice message, not a real customer. I'll update this note once I've used it on real ones.
The rule I'd keep
The first reply doesn't have to fix everything. It has to show you've understood and say clearly what happens next. Where the AI isn't sure, it should write [CHECK], not guess.
Before you hit send
- It names their actual problem, not a generic "issue".
- One apology at most, for something real.
- Every [CHECK] is answered, with a time you can keep.
- You'd be comfortable if they shared it.
Already a public review? See replying to a bad review.
If this was useful, these might be too
A note in your inbox on Tuesdays
Once a week I send one task, the prompt I use for it, and anything I got wrong along the way. It takes two minutes to read.
Free. Unsubscribe whenever you like.