Difficult Emails: Complaints, Apologies and Saying No
These are the emails people put off the longest: telling a client you made a mistake, pushing back on an unreasonable request, delivering a complaint about someone else’s work, or simply saying no to something you cannot or will not do. The stakes feel higher because the relationship is genuinely on the line, and that pressure is exactly what makes people freeze, overwrite, or send something too defensive or too apologetic. AI is particularly valuable here, not because it feels less anxious than you do, but because it can produce a calm, well-structured first draft that you can then adjust, giving you something to react to and edit rather than a blank page to fill under stress.
Apologies: own it once, then move forward
The most common mistake in AI-assisted apology emails, and in human-written ones, is over-apologising, which reads as anxious rather than accountable, and paradoxically undermines the reader’s confidence that you have things under control. A strong apology email states clearly what went wrong, takes ownership without excessive self-flagellation, explains what is being done about it, and states a concrete next step, all without repeating “sorry” more than once or twice.
A strong prompt: “Write an apology email to a client. We shipped the wrong product variant to them last week. Acknowledge the mistake clearly, once, no excessive apologising. Explain we’ve identified it was a warehouse labelling error and have already fixed the process. State that the correct item ships today with next-day delivery, and offer a 10 percent discount on this order as a gesture. Confident and accountable tone, not groveling. Under 130 words.” The instruction “not groveling” is doing real work here, since AI models, if not told otherwise, sometimes over-apologise by default, stacking multiple “I’m so sorry” phrases that make the writer sound less in control, not more sincere.
Handling complaints, whether you are sending or receiving one
When you are the one raising a complaint, the goal is to state the problem factually, avoid inflammatory language even if you are frustrated, and be specific about the resolution you want. A prompt like “Write a complaint email about a delayed delivery, three weeks late with no communication from the supplier. State the facts and dates plainly, no angry language even though I am frustrated, and ask specifically for a revised delivery date within 48 hours plus a partial refund for the delay” keeps the email professional and effective, since even a legitimate complaint written in anger tends to get a defensive rather than cooperative response.
When you are receiving and responding to a complaint, the priority shifts to acknowledging the person’s frustration before addressing the facts, since people who feel unheard tend to escalate. Ask the AI explicitly for this order: “Write a reply to an angry customer complaint. First, acknowledge their frustration genuinely, don’t just say I understand, name the specific problem they raised. Then explain what happened factually. Then offer the resolution: a full refund processed today. Calm, empathetic, no defensiveness, don’t blame the customer for anything even if they made an error.”
Saying no without damaging the relationship
| Part | Purpose | Example phrasing to request from the AI |
|---|---|---|
| Brief acknowledgement | Shows you took the request seriously | “Thanks for laying out the request in detail” |
| Clear decision | States the no directly, without excessive hedging | “We’re not able to accommodate this for the current phase” |
| Brief reason | Gives enough context without over-justifying | “Given the timeline and budget already agreed” |
| Constructive alternative | Leaves the relationship in good shape | “We could revisit this for phase two, or explore X instead” |
A prompt built on this structure: “Write an email declining a client’s request to add two extra features to the current project scope without additional budget or time. Be clear and firm, one brief reason, no excessive apologising or over-explaining, then offer to include these in a phase two proposal instead. Professional, not cold.” The instruction “no excessive over-explaining” matters because AI drafts, like many people, tend to pad a refusal with justification, which can read as defensive and actually invites more pushback than a clean, confident no.
Checking the emotional accuracy of the draft
Difficult emails are the category where you must read the AI draft most carefully before sending, because tone missteps here carry the highest cost. After generating a draft, ask the AI directly to critique it: “Read this apology email back and tell me if any part of it could come across as defensive, dismissive, or insincere.” This kind of self-critique prompt, treating the AI as a second reader rather than only a drafter, catches problems you might miss when you are too close to the situation, and it is worth using on every difficult email before sending, regardless of how confident the first draft feels.
What AI cannot do here
AI does not know your organisation’s policy on refunds, discounts, or liability, and it cannot make a commitment on your behalf that you are not authorised to make. Never let an AI-generated draft commit to a specific remedy, a refund amount, a compensation gesture, a policy exception, that you have not already confirmed you are able to offer. Treat the AI as a drafting partner for tone and structure, never as the source of what you are actually promising.
Difficult one-to-one emails require care and precision. The next lesson shifts to a different register entirely: newsletters and marketing emails, written for a broad audience rather than a single reader, where AI’s strengths in structure and variation get used in a different way.
