How to Write a Difficult Message With AI (Bad News, Pushback, Apologies)

AI is genuinely useful for difficult messages, but not the way most people use it. Its job is not to write the bad news for you; it is to help you structure the message, strip the accidental harshness or mush out of your draft, and pressure test how it will land before you hit send. The words that finally go out should be yours, because the relationship on the other end is yours.

What makes a difficult message go wrong?

Burying the point and blaming the reader. Most bad-news messages fail in one of two ways: three paragraphs of throat clearing before the actual news, or blunt delivery with no context and no path forward. Research summarized by Harvard Business Review on delivering bad news points the same direction: people accept hard news far better when it comes early, with honest reasons, and with clarity about what happens next. Vague softening does not spare feelings; it creates a second, worse conversation later.

How should I structure bad news?

Context, the news itself, then next steps. One or two sentences of context, the news stated plainly in the first third of the message, the reason honestly given, and concrete next steps with owners and dates. Prompt: 'Help me structure this message: [situation]. One short context sentence, the decision stated clearly, the honest reason, then next steps. No corporate padding, no false cheerfulness.' Communication coaches like Science of People offer script patterns built on the same skeleton.

How do I fix the tone without losing the message?

Draft honestly, then ask for a tone pass. Write the ugly first version yourself, anger and all, then ask: 'Keep every fact and decision in this draft, but remove blame language, sarcasm, and hedging. Make it direct and respectful.' The AI is unemotional, which is exactly what you need at this step: it will not escalate, and it will not water down the decision to feel nicer. Compare its version with yours and keep the sentences that still sound like you; our guide on training AI to sound like you helps with that last part.

How do I know how the message will land?

Make the AI read it as the recipient. Prompt: 'Read this as the person receiving it, who expected a different outcome. What stings unnecessarily? What questions will they immediately have? What is unclear about what happens next?' Fix what it finds, then decide the channel honestly: serious personal news deserves a conversation, with the written message as follow-up. Email works for decisions that affect process; it is a poor place for news that affects a person. For high-stakes messages in a language you did not grow up with, AI is also a fairness tool, separating what you mean from how fluently you can phrase it.

When should I wait before sending?

Whenever the message was written hot. If you drafted angry, park it for an hour, or overnight for anything career-shaped, then reread the AI-polished version once more before sending. Two final rules: never send a difficult message you have not fully read aloud at least once, and never outsource the decision itself to the tool. AI can shape how you say it; whether and what you decide stays human.

Clear is kind. The message that hurts most is the one the reader has to decode.

Next time a hard message is sitting in your drafts, run the loop: structure it, tone pass, recipient read, cool down, send. Ten extra minutes, and you only have to have the conversation once.