Using AI for work emails without sounding like a robot.

Everyone has now received one: the email that opens "I hope this message finds you well", thanks you for your valuable input, and closes with warm regards from a colleague who has never once used those words. AI did not write a bad email. It wrote a generic one, and generic is what gets noticed.

Used properly, an AI assistant is the best email tool since the search bar - not because it writes for you, but because it untangles threads, breaks blank pages and takes the emotional heat out of difficult replies. The craft is in keeping your own voice. This works the same in ChatGPT, Copilot, Gemini or Claude; Copilot and Gemini simply sit closer to your inbox.

Why do AI emails sound robotic, and what fixes it?

Assistants are trained on oceans of corporate text, so their default register is an average of every office email ever written: over-polite, over-hedged, faintly American. Ask for "a professional email" and that average is exactly what you get. The fix is to stop describing tone and start showing it:

Three past emails beat any adjective. Save this prompt with your examples baked in - in a notes file you paste from, in your assistant's custom instructions, or in a saved project if it offers them - and it becomes a one-line tool: "reply to this, usual style". Save it somewhere, though - memory features are inconsistent between tools and chats, and you do not want to depend on the assistant having kept last week's brief. The short version only works reliably once the examples are in front of it again. The general pattern - context, task, material, standards - is unpacked in our starting-point guide.

A before and after.

The task: chase a colleague for figures that are a week late, second reminder, without burning the relationship. The assistant's default attempt:

"I hope this email finds you well. I am writing to gently follow up on my previous request regarding the Q2 figures. I completely understand how busy things can get! When you have a moment, I would be incredibly grateful if you could share an update at your earliest convenience."

Fifty words of apology for existing, and not one of them is the actual request: no date, no reason, nothing the reader can act on. The same request with the tone anchor and one added instruction - "be direct and friendly, name the deadline and the reason for it":

"Quick nudge on the Q2 figures - I need them by Thursday to get the board pack out on Friday. If something's blocking you, tell me and I'll see if I can help. Otherwise, Thursday morning would be perfect."

Thirty-eight words, sounds human, states the deadline and the consequence, offers help. Two minutes of briefing against what might otherwise be fifteen minutes of drafting and second-guessing - and this is a message most people put off for a day. The time saved on email is usually less about typing and more about not procrastinating, a pattern that repeats across the whole working week.

Which difficult emails are worth briefing properly?

• The chase: as above - name the deadline, the reason and an offer of help. Ask the assistant for "firm but warm, no apologising for asking".
• The no: "Draft a reply declining this request. Give the real reason in one sentence, offer one genuine alternative if there is one, no fake regret." A clean no reads better than a padded maybe.
• The apology: "Acknowledge the specific mistake, say what we are doing about it and by when, no excuses, no 'any inconvenience caused'." Then check every commitment in the draft is one you can keep.
• The ask: for favours or upward requests: "Make the request in the first two lines, give the context after, make it easy to say yes with one clear question."

For all four, the assistant's real value is emotional distance. It is not annoyed, embarrassed or defensive, so its draft starts from neutral - and you add the humanity back in the edit.

Use it on the inbox, not just the outbox.

Drafting is half the win. The other half is reading. Paste a 30-message thread and ask: "What is being asked of me, what has already been agreed, and what would I contradict if I said yes?" Paste the unread pile from a week off and ask for a triage list: needs a reply, needs an action, information only. Summarising is lower risk than drafting because errors are private - though the checking rule still applies before you act on a date or figure the assistant extracted. The same prompt structures appear in our reusable work prompts.

Which emails should AI never write?

Some messages carry their value in the fact that you wrote them. Condolences. Congratulations that matter. Apologies for serious failures. References. Anything in a live conflict, grievance or performance process - where a misjudged sentence has consequences and "the AI drafted it" is not a defence. If the recipient would be hurt to learn a machine wrote it, that is your answer. And before pasting any thread into a public tool, apply the confidentiality test: no client data, no personal information about colleagues, no unreleased numbers. The full boundary rules are in using AI safely at work.

Frequently asked questions