Skip to content

Type something to search the manual

    Hard emails

    ~ min read

    30-second summary
    • The emails you dash off aren’t the problem. The hard ones are, and that’s where AI needs a different role: not ghostwriter, but a consultant on tone.
    • Three situations where it earns its keep: saying no without burning bridges, responding to someone who messed up without punishing, giving delicate feedback without judging the person.
    • The three-step pattern: describe the full situation (don’t just paste the email), ask for versions with different registers, compose your own from pieces.
    • The last hand is yours. The signature too.

    A teammate has missed an important deadline: you have to tell them without punishing, but without softening what happened. A client keeps pressing for a discount you can’t give: you say no and you’d still like to keep the relationship. A peer, in copy to ten people, has put you publicly on the spot: you answer today, and everyone reading registers how you react.

    Normal emails you write without thinking, and using AI out of the box is fine for those. These are something else. Here the wrong word costs a relationship, a client, or months of good work gone to waste.

    For these emails AI shifts role. It’s no longer the ghostwriter producing a text you paste. It’s a consultant on tone: you describe the situation, ask for versions from different angles, and only take pieces. The last hand stays yours.

    Three situations where AI really earns its keep

    Section titled “Three situations where AI really earns its keep”

    No’s are the most frequent case: a client who wants a discount beyond your margins, a colleague who asks you to take on a project when your plate is already full, an invitation to speak at an event that doesn’t fit. A bare no rarely works: it leaves the other person to guess the reason on their own, and without one they imagine something worse than the reality.

    The pattern that works well is “no + reason + bridge.” No need for extended apologies: a crisp reason and, where possible, a future opening (“not now, I’ll check in after June” / “not on this, but on X yes”). The AI produces good variants of this schema with different registers, and it helps you find the point where your opening sounds sincere and not perfunctory.

    A concrete example of a refusal to a client asking for a discount beyond margins: “Thanks, and no: below this threshold production barely breaks even and quality would suffer. If you have flexibility on timing (one more month) we can fit your original budget, otherwise we stay at the current price.” Three moves: clear no, operational reason, concrete opening. If the bridge isn’t there (a client you won’t re-engage, a request that won’t come back), say so without forcing fake alternatives: “…otherwise we stay at the current price, no hard feelings.”

    A supplier delivers late, a colleague hands you shoddy material, a client is trying to shift the blame to you for their own mistake. You need a firm tone that doesn’t punish, and above all one that doesn’t get you labeled “difficult.”

    The dominant risk is writing on impulse while you’re still angry. Handing your draft to the AI acts as a filter: you ask it to read and tell you if it sounds hostile, passive-aggressive, or if you’re loading a tone you don’t mean. A prompt that works well: “Read this draft. Does it sound hostile or passive-aggressive? Does it feel like I’m punishing instead of asking for clarity? Flag the spots where the tone isn’t what I wanted.” Often that’s enough to make you rewrite two or three sentences, and the email comes out different. If the stakes are public (many people in copy), the filter is worth twice as much.

    This situation and the previous one overlap, but they aren’t the same case: there you’re responding to something the other side just did that pulls you in; here you’re the one opening the conversation, with no immediate trigger, because something needs to be said. A junior whose work isn’t up to standard, a peer who made a call you think needs repairing, a boss who’s been managing you badly. In all three cases the content has to land clearly, but the person on the other side shouldn’t feel an attack on their identity.

    The rule that guides these emails is to state the fact without judging the person. Write as if the email might be read by a stranger: would it be clear what happened, without unnecessary attributes on the person? If yes, you’re fine. If not, there are surplus adjectives to cut. The AI helps here as a mirror: you ask it to flag the passages where you’re evaluating instead of describing, and it marks the sentences.

    1. Describe the situation, not just the email

    Section titled “1. Describe the situation, not just the email”

    The difference between an email that sounds right and one that sounds like a business manual is all here. Pasting the draft and saying “improve it” produces the AI-polish you encountered in Write better: polite, smooth, unsigned.

    Instead you describe: who this person is (role, how long you’ve worked together), how your relationship usually goes (formal, casual, tense), what happened before this email, what you want to achieve with the reply, what you absolutely don’t want (“I don’t want it to sound like I’m punishing them, I don’t want them to feel they can keep postponing”). The more you tell, the better the AI reads the emotional temperature of the message.

    If the stakes are urgent (you have to send within an hour), compress the description into three or four essential sentences: who, what happened, what to avoid. The pattern still holds, it just loses depth.

    2. Ask for versions with different registers

    Section titled “2. Ask for versions with different registers”

    At this point the Write better pattern applies, with finer labels: “Give me three versions: one more direct, one more diplomatic, one more terse. Four sentences max each.”

    Strong labels force the AI to move the register. “One normal” produces nothing useful: it produces the default. “One more terse,” “one less formal,” “one with a personal opening,” move the text in recognizable directions.

    3. Compose your own, don’t take a whole one

    Section titled “3. Compose your own, don’t take a whole one”

    None of the versions is the answer. They’re three angles of attack, and your final email is usually a mosaic: opening from the first, body from the second, closing of your own. Then you read it out loud one last time.

    If reading out loud you find it different from how you usually talk, you’ve left too much AI voice. Go back and trim.

    A junior teammate of yours has missed an important deadline with a client: the document was due yesterday, it went out today. You two have worked together for eight months, she’s usually on time, this is her first serious slip. You don’t want the email to sound punitive (it’s an isolated case) but you don’t want to play it down either (with the client, credibility took a hit). When you write the prompt, replace the real name before pasting. Three ways that work: a generic (“the teammate”), a fake name (“Clara” in place of the real one), or just initials (“C.”). The AI has all the context it needs without getting the actual name.

    You
    I need to write to my junior teammate about a document delivered to the client yesterday instead of the day before. We've worked together for 8 months, she's usually on time, this is her first serious slip. I don't want to sound punitive because it's an isolated case, but I don't want to play it down either. Give me three versions: one direct, one diplomatic, one terse. Four sentences max each.
    AI
    Direct version Hi [name], the document for [client] arrived today instead of yesterday and the client noticed this morning. It happens, and with you it never does: that's why I'd like to understand what changed this week. Shall we take 20 minutes tomorrow morning? I'll take care of the client. Diplomatic version Hi [name], I'm writing about something that came in this morning. The client noticed that the document they expected yesterday went out today, and they wrote to me. Since it's the first time this has happened, I wanted to look into it with you and see what slowed the delivery, so we can line up alerts for the next deadlines. Coffee tomorrow at 10? Terse version Hi [name], the document for [client] went out a day late and the client noticed. I'll handle the client, but the delivery was yours and I'd like to know what happened. I'm booking 20 minutes tomorrow at 10, small meeting room.

    The three registers aren’t obvious: the “direct” is warm (“with you it never does”) and shifts the focus to understanding; the “diplomatic” is colder than it seems, because the “coffee” softens but the frame stays formal; the “terse” is the most functional but risks sounding stiff if your relationship with her is warm. Your final version probably steals the opening from the first, the reason from the second, and the practicality from the third. You put [name] and [client] back in yourself just before sending: the AI never saw them.

    Don’t delegate the whole drafting. On hard emails the content is the tone, and the tone comes from your relationship with the person. Pasting the AI version without reading it is what leads to signing texts that aren’t yours. The corollary is in Things NOT to do: in certain situations the decision is yours, even if someone could wrap it up for you the same way.

    Don’t send the first version. Read out loud. If the sentence reads smoothly, move on. If you stumble, change it. If you feel a bit embarrassed saying it out loud, it’s one you’d never write: it’s the AI’s.

    Don’t paste third-party data without anonymizing. The name of the person receiving the email, the client, the figures at stake: strip them or replace them before handing the text to the AI. A generic works (“[name]”, “the colleague”), so does a fake name, or just initials; what matters is that the AI doesn’t see the real data. The privacy of the people involved in your email isn’t yours to give away. The full picture is in What you share when you use AI.

    Hard emails are a specific case of high-stakes writing. From here on the module moves from short messages to long documents: proposals, briefs, reports. The next lesson, Drafting professional documents, opens that chapter.