How to Write Difficult Messages With AI: Refusals, Collections, Apologies

How to Write Difficult Messages With AI: Refusals, Collections, Apologies

To write a difficult message with AI, tell the model who is sending it, who is receiving it, which fact does not change and what outcome you want, then ask for a short, firm and respectful draft. After that, review everything calmly. AI helps organize the words, but the person who sends the message remains responsible for what was said, promised and refused.

Key takeaways

• A difficult message improves when your request to the AI includes facts, limits and a goal, not just an instruction to be polite.
• Refusals, payment reminders and apologies have different structures, and each needs its own kind of instruction.
• Personal and sensitive data does not need to go into the tool: use placeholders for names and amounts until the final version.
• Human review is the step that guarantees tone, accuracy and commitment. The AI proposes, the person decides.
• In teams, standardize the request and the review checklist instead of relying on each person's talent with prompts.

Why difficult messages need more care than ordinary ones

A difficult message is one where the text can change the relationship. When you decline a request, chase someone who is late or admit a mistake, the person on the other end reads every word looking for intent. One extra adjective sounds like irony, a vague sentence sounds evasive, and an apology that runs too long sounds like an excuse.

Generative AI is good at producing fluent, polite text, and that is exactly the risk. The result often comes out sounding like a press release: correct, but distant. It can also invent a deadline, promise something the company has not decided, or soften a refusal so much that the customer concludes there is still a chance.

The right way to use it is to treat the model as a fast writer who does not know the situation. It does not know what was agreed, what the company can deliver or how this particular person tends to react. You do. The work of writing the message well starts before the prompt and ends after the answer. To decide what you should never hand to the model, also read what not to delegate to AI in your business.

What to gather before asking for a draft

The quality of the draft depends on the material you provide. Before opening the tool, write down these items in a few lines:

• Who is sending and in what role (owner, manager, support, finance).
• Who is receiving and what the relationship is (long-time customer, supplier, employee, partner).
• The fact that does not change, written without adjectives: what was agreed and what happened.
• The decision already made, or the action you need the other person to take.
• What you can offer, if there is an alternative, and what you cannot promise.
• The channel and the length: WhatsApp calls for short text, email can hold a bit more.
• The tone you want, with an example sentence you would say in person.

The last item makes a difference. When you paste two or three sentences of your own, the model imitates your voice instead of falling back on generic customer service language. If the message will go out on WhatsApp, remember that company-initiated messages have their own rules, covered in WhatsApp outbound messages: templates, opt-in and best practices.

How to ask for a firm, respectful refusal

A good refusal follows a simple order: acknowledge the request, state the decision early, explain the reason in one sentence, offer what is possible and close without opening a negotiation you do not want. AI tends to start with many apologies and leave the decision for the end. Ask for the opposite explicitly.

A request that works well looks like this:

Write a refusal of no more than five sentences, in a cordial and direct tone: state the decision in the first sentence, explain the reason in one sentence, offer the alternative I give you and promise nothing beyond that.

After the request, supply the facts and the alternative. If there is no alternative, tell the model so it does not invent one. A common mistake is asking for a kind refusal and letting the model decide how much to concede. It concedes too much, because it was trained to please.

A good practice is to ask for two versions: a shorter one and a slightly warmer one. You compare and choose, which also helps you notice which one sounds like you.

How to write a payment reminder without threats or beating around the bush

Collecting well means saying what is open, since when and what needs to happen now, without sounding accusatory. The model tends to slip toward two extremes: a text so soft that nobody realizes there is an outstanding item, or a text so dry it reads like a formal notice. Ask for a clear structure.

For a payment reminder, this plan works as a step by step:

• Open with a sentence of context, without blame: you are following up on something that was agreed.
• Describe what is pending, referring to the document or the agreed date, using placeholders in the draft and filling in the real data only at the end.
• State which action you expect and by when.
• Offer a path for people who are struggling, such as a conversation about how to settle the balance.
• Close with a contact channel, with no threats or consequences the company has not decided to apply.

Handle any mention of credit bureau reporting, protest of a debt or legal action with care. That kind of sentence has legal and reputational effects, and the company decides whether it belongs in the message, with legal support when needed. Tell the AI not to include any consequence you have not provided.

On data: the draft does not need the real name, tax ID or exact amount. Use placeholders such as [CLIENT], [AMOUNT] and [DATE], and replace them afterward. This reduces exposure of personal data and supports the data minimization principle in Brazil's data protection law (LGPD). If your company sends data to tools hosted outside the country, see international data transfer under LGPD when using AI providers.

How to apologize without overdoing it or getting defensive

An apology is the easiest message to spoil with AI, because the default language model text sounds rehearsed. Phrases like we deeply regret any inconvenience this may have caused say little and do not acknowledge the mistake. A good apology has four parts: say what happened, take responsibility without hiding behind third parties, explain what will be done and say when the other person will hear the result.

Tell the model to avoid three things: long justifications, public relations language and any commitment you have not confirmed. If the error is still being investigated, say so in the request, so the text acknowledges the problem without stating the cause.

Illustrative scenario: a services company delayed an agreed delivery and the customer complained on WhatsApp. The person in charge gives the AI the facts (agreed deadline, what happened, the new estimate confirmed by the team) and asks for a text of up to five sentences. The first draft comes back sounding like an announcement and promises delivery by the end of the week, something nobody confirmed. In review, the person replaces the promise with the real estimate, cuts the sentence about deep regret and adds one line saying when they will follow up. The final text is shorter, more truthful and in the voice of the person actually sending it.

When the channel is public, such as consumer complaint sites, the care needed is greater, because the text is also read by people who are not yet customers. The article handling consumer complaint sites with AI: draft, review, send covers that case in detail.

Review checklist: what only a person can verify

Review is the part that keeps the message human and keeps responsibility with the sender. Before sending, check point by point:

• The facts are correct: dates, amounts, names, what was actually agreed.
• There is no new promise the company has not confirmed.
• The decision appears early and without ambiguity, especially in a refusal.
• The tone is yours: would you say this sentence out loud to this person?
• No sentence sounds like a threat, irony or blame placed on the other person.
• There is no unnecessary personal data, and all placeholders have been replaced.
• The text respects the person's context: a customer in a hard situation, an employee hearing the news for the first time.
• The recipient knows the next step and who to talk to.

For messages that affect employment, a contract, health or reputation, add a second person to the review. In the most sensitive cases, such as a dismissal or a conflict between partners, the conversation should happen by voice or in person, and the written message only confirms what was already said. AI can help organize what you will say, but it does not replace the human moment.

How to turn this into a team routine

If several people in the company write difficult messages, the gain comes from standardizing the process, not just the prompt. The method we use in AI projects applies with few adjustments: map, limit, pilot and hand over.

• Map: list the types of difficult messages your team really sends (declining a request, payment reminders, correcting a mistake, delay notices) and who sends them today.
• Limit: define what can go into the tool (facts without personal data) and what cannot (documents, sensitive data, health information), plus what the AI never decides, such as discounts, deadlines or legal action.
• Pilot: pick one type of message, use the request template and the checklist for a short period, keep the approved drafts and note what had to be corrected.
• Hand over: based on what worked, write the final request template, give it to the team and decide whether to expand to other message types, adjust what failed or stop.

If your company does not yet have rules for tools like ChatGPT, start with a company policy for employee ChatGPT use. It defines what can be pasted in, who reviews and where records are kept. To structure reusable instructions, it also helps to read how to write instructions for an AI support agent.

A note on sources: technical risk management guidance, such as the NIST AI risk management framework, argues for human oversight proportional to the impact of the decision. It is a good practice reference, not a rule specific to messages. Applying it to refusals, payment reminders and apologies is our own interpretation, based on implementation experience, and should be adapted to your context.

Frequently asked questions

Can AI send a difficult message on its own? That is not recommended. In messages involving a refusal, a payment reminder or an error, the text has consequences for the relationship and sometimes for the contract. Use AI to draft and a person to approve and send.

Can I paste the customer's name and data into the prompt? The safest approach is to use placeholders in place of the name, ID number and amounts, and fill them in only in the final version. That reduces exposure of personal data and makes LGPD compliance easier. Also check your company's internal rules and the tool's terms.

How do I stop the text from sounding like a robot? Give the model two or three sentences you would write yourself and ask for short text without announcement formulas. Then read it out loud and cut everything you would not say to that person. The final review matters more than the perfect prompt.

Is it worth creating a prompt template for the team? Yes, as long as it includes the required facts, the limits on what can be promised and the review checklist. Start with one type of message, watch what the team has to correct and only then expand. This keeps each person from inventing their own standard.

Sources

• NIST: AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados
• Google Search Central: helpful, reliable, people-first content