AI for a Small BusinessLesson 33 of 88

Contents

AI for a Small Business

Lesson 33 of 88. Sign in to keep your place.

Module 4 · Customers, from first message to resolved complaint

The angry message, and what your reply commits you to

9 min read

Where a drafting tool helps

It is nine in the evening and somebody has written something unfair about work you were proud of. You are not the right person to write the first draft.

A machine that is not angry, has not been insulted, and does not need to win produces a calmer opening than most people manage in that state. This is one of the clearest wins in the course, and it is worth having even if you rewrite every sentence, because the draft breaks the block, and a blocked complaint reply gets answered on Thursday, which makes everything worse.

Two failure modes, both predictable

It over-apologises. The default output accepts fault before anyone has established there was any. “We’re so sorry we let you down” is a sentence about your liability, not about your manners, and you may not yet know whether the parcel was late or the customer gave the wrong address.

It over-promises. Refunds, replacements, free redelivery, a discount on the next order. Agreeable text is what these systems were tuned toward, and generosity reads as agreeable. The model has no idea that this margin is 8% and that this customer has done the same thing three times.

Both come from the same place and neither is fixable by asking for a better tone.

The commitment audit

One pass, twenty seconds, before any complaint reply goes out. Read only for commitments and ignore everything else.

Ask of each sentence: what does this oblige me to do? Then delete anything you did not decide to give. Not soften: delete. A promise softened is still a promise.

Watch for the phrasings that commit without looking like it: “we’ll make sure this doesn’t happen again”, “we’ll get that sorted for you today”, “of course we’ll cover the return postage”. Each is a specific undertaking dressed as sympathy.

Apology and admission are not the same

Worth understanding rather than guessing at.

An expression of sympathy (“I’m sorry this happened, that must have been frustrating”) is different from an admission of fault: “I’m sorry we sent the wrong item.” Several jurisdictions have gone further and legislated that an apology is not by itself an admission of liability, precisely because people were staying silent to protect their position.

The law varies and you should not take a position from a course. The practical craft, which is safe nearly everywhere, is:

  • Sympathy, freely. It costs nothing and it de-escalates.
  • Facts, only the ones you have checked.
  • Fault, only when you have established it.
  • Remedy, only what you decided.

Put it in your prompt as a standing constraint:

Text
Do not apologise for anything not stated as a fact below.
Do not offer refunds, replacements, discounts or postage.
Acknowledge the feeling, state what I will do next, give
a time. Nothing else.

The ladder

Most complaints resolve in three moves, and having them written down means you do not invent a policy at nine in the evening.

  1. Acknowledge and find out. Sympathy, no fault, one specific question, a time by which you will come back.
  2. Decide and say so. Now you know what happened. State the remedy plainly, once. Do not re-apologise; repeated apology reads as weakness and invites escalation.
  3. Close. Confirm what was done, in writing, with a date. This is the message that matters if it ever goes further.

Draft all three. Read all three. Send them yourself.

Never automatic

Complaint replies do not get auto-sent, at any volume, on any tier. The category exists precisely because something has gone wrong, and the cost of getting the reply wrong is not one lost sale: it is a public review, a chargeback, or a regulator.

Draft, always. Send, never.

Three questions before you draft anything

Answer these yourself, in ten seconds, before opening the tool. They are what the model cannot know and what determines the whole reply.

Is the customer right? Sometimes obviously yes, sometimes not yet known. The reply differs completely, and the model will assume yes.

What is this worth? A regular customer of six years and a first-time buyer of a low-value item are different commercial situations. Nothing about the message tells the model which one this is.

What am I willing to give? Decide before drafting. Deciding while reading a warm, apologetic draft that already offers a replacement is not deciding.

Put the three answers into the facts section of the prompt. The draft that comes back will then be about your situation rather than about complaints in general.

When it goes public

A public review or a comment thread is a different act with a different audience: not the complainant, but everybody reading later.

Keep it short, factual, unemotional, and move it off the platform quickly: acknowledge, state the one relevant fact, invite them to contact you directly. Long public replies always read badly, and a model asked to be thorough will write one. Give it a word limit of about sixty and delete the last line.

Every lesson here is open. Your progress is recorded once you pass the AI at Work exam.

Check your understanding

An owner adds “be warm and empathetic” to her complaint-reply prompt. Drafts improve in tone and she notices they now frequently offer free replacements. What is the connection?

Pick the one you would defend. Your answer stays private.

Next lesson · 9 minBookings and no-shows: the case where AI is the wrong tool

Stuck on something?Ask Yetunde, your AI tutor

How to cite: Addaly (2026) ‘The angry message, and what your reply commits you to’, in AI for a Small Business. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-small-business/complaints-and-refunds (Accessed: 13 October 2026).