AI for email and writing: draft faster without sounding like a robot

Where AI helps with writing, and where it hurts

AI is very good at three writing jobs: producing a first draft from rough notes, cutting a long message down, and pulling the thread of a 14-message email chain into a reply. It is bad at one thing that matters: sounding like you, unless you show it how.

The robot voice you have seen in your inbox ("I hope this email finds you well! I wanted to reach out regarding...") is not the AI's fault. It is the default output of a prompt that said "write a professional email." The fix is not a better adjective. It is giving the model your actual writing to copy.

This guide covers how to do that, the five everyday writing jobs with a prompt for each, what good and bad output look like side by side, the AI already built into Gmail and Outlook, and the etiquette rules that keep you out of trouble.

Give it your voice first

Spend fifteen minutes once and every draft after that gets better.

Build a voice sample

Open your sent folder. Find three to five emails you were happy with: one routine update, one that said no, one that asked for something, one to a client or a senior person. Copy them into a note, strip out names and anything confidential, and label each with who it was to and why it worked.

That note is your voice sample. Paste it into any prompt where voice matters, or save it as the instructions in a custom assistant (a Custom GPT, a Claude Project, or a Gemini Gem) so it is always there.

Write the voice note

Under the samples, add a short list of rules about how you write. Be specific and negative where you can, because "avoid X" is easier for a model to follow than "be more Y." An example:

• Greeting is "Hi [FIRST NAME]," with a comma. Sign-off is "Thanks," then my name.
• First sentence says what I need or what changed. No warm-up lines.
• Short paragraphs, two or three sentences. Bullets only when there are three or more items.
• Never use "I hope this finds you well," "reach out," "circle back," "touch base," or "per my last email."
• No exclamation marks in messages to clients. One is fine with colleagues I know.
• If I am asking for something, the ask is in its own sentence with a date.

With the samples and the note, "write this in my voice" becomes a real instruction instead of a wish.

The five jobs

Draft from bullet points

The fastest way to write an email with AI is to not write the email. Write the bullets, then let the model do the connective tissue.

You are drafting an email that I will send from my own account, in my voice. Match the voice samples and the rules below exactly.

Voice samples:
[PASTE 2 TO 3 EMAILS YOU SENT]

Rules:
[PASTE YOUR VOICE NOTE]

Recipient: [WHO THEY ARE, HOW WELL I KNOW THEM, WHAT THEY ALREADY KNOW]
Purpose: [WHAT I WANT THEM TO DO OR KNOW AFTER READING]
Points to cover, in order:
- [POINT 1]
- [POINT 2]
- [POINT 3]

Length: [UNDER 120 WORDS OR UNDER 200 WORDS]. Do not add facts, dates, or commitments that are not in my points. If something seems missing, add a line at the bottom starting with "QUESTION FOR ME:" instead of guessing.

The last sentence is the important one. AI drafts fill gaps with plausible details, like a delivery date you never promised. Telling it to ask instead of guess is the difference between a draft you can trust and one you have to reverse-engineer.

Reply to a long thread

Paste the whole thread and ask for two things: what is actually being asked of you, and a reply.

Below is an email thread. I am [YOUR NAME, YOUR ROLE]. First, in three bullets, tell me: what is being asked of me, by whom, and by when. List any open questions the thread has not resolved. Then draft a reply in my voice (samples and rules below) that answers only what I have decided, which is: [YOUR DECISION OR ANSWER]. Under 150 words. Do not commit me to anything beyond that.

Thread:
[PASTE THE EMAIL THREAD]

Voice samples and rules:
[PASTE]

Read the bullets before you read the draft. If the model misread the thread, the draft is wrong no matter how good it sounds.

Rewrite for a different reader

Same content, different audience: an update written for your team, rewritten for the executive who reads the first two lines and nothing else. The prompt is short: "Rewrite this for [READER]. They care about [WHAT THEY CARE ABOUT] and will not read past the second paragraph. Keep every fact; change nothing that is a number or a date."

Tighten

Paste the long version and say: "Cut this to half the length without losing any decision, date, name, or number. Return the cut version, then a list of what you removed." The removal list is your safety check. Scan it for anything that should go back in.

Match tone

Tone requests work when they point at something concrete: "warmer, like my email to Dana below," "firmer, this is the third reminder," "less apologetic; we did nothing wrong." They fail when they use words like "professional," which the model reads as permission to be stiff.

Good output and bad output, side by side

The task: tell a client their project will be one week late because a subcontractor missed a delivery.

A bad draft, from the prompt "write a professional email to a client about a delay":

I hope this email finds you well. I wanted to reach out to inform you that, due to unforeseen circumstances beyond our control, we are experiencing a slight delay in the project timeline. We sincerely apologize for any inconvenience this may cause and appreciate your understanding and continued partnership. Please don't hesitate to reach out with any questions.

Nothing in it is false, and nothing in it is useful. It hides the length of the delay, blames no one and everyone, and offers no plan.

A good draft, from a prompt that included the voice sample, the facts (one week, subcontractor delivery, the new date, what is being done), and the rule "lead with the change":

Hi Marcus,

The install is moving from the 14th to the 21st. Our electrical sub missed their delivery window on the panels, and I would rather tell you now than hope they catch up.

What I am doing about it: the panels are confirmed for the 16th, and I have moved the crew so we lose no further time once they land. The 21st is a date I am confident in.

If the 21st causes a problem on your side, call me today and we will work it out.

Thanks, Jordan

The difference is not intelligence. The second prompt contained the facts, the reader, and a voice. The first contained an adjective.

The AI already in your inbox

You may not need a separate chat tool for most email work.

Gmail with Gemini can draft from a prompt inside the compose window, offer rewrite options (formalize, elaborate, shorten, polish), and summarize a long thread in the side panel. It sees the thread you have open, which saves the paste step.

Outlook with Microsoft 365 Copilot does the same jobs: Draft with Copilot in a new message, a coaching feature that critiques tone and clarity before you send, and thread summaries at the top of long conversations. Availability depends on your organization's license.

Grammarly sits on top of whatever you type, flags tone (it will tell you a message reads as blunt or uncertain), and offers rewrites. It is good for the last pass, less good for drafting from nothing.

These built-in tools have one advantage that matters for confidential work: they run inside your organization's approved environment, so a thread with client details stays where your IT and legal teams already expect it. Check your policy, but the built-in tool is usually the safer default. See AI privacy at work for the full picture.

Workplace etiquette

Every draft gets read in full

Not skimmed. Read. AI drafts have three predictable failure modes: they invent a detail to fill a gap, they soften a commitment ("we should be able to") or harden one ("we will"), and they pad. Any of the three can cost you. If you would not send an intern's draft unread, do not send the model's.

Disclosure

For routine email, nobody discloses that spell-check helped, and drafting help falls in the same category, as long as the content, the decision, and the responsibility are yours. Three exceptions. If your organization has a policy, follow it. If the recipient asks, answer honestly. And if the message is personal in nature (condolences, a thank-you for something that mattered, a reference), write it yourself, because the reader can usually tell, and because writing it is the point.

What should not go through AI at all

Anything where the wording carries legal or HR weight: performance reviews, termination or discipline notices, complaints, anything about an employee's health, compensation, or a dispute. Draft those with the people whose job it is, in the systems built for it. The same goes for regulated content: patient information (HIPAA), client matters covered by attorney-client privilege, and personal data (PII). Never paste confidential data into a consumer AI tool unless your organization has approved it.

Do not reply faster than you think

AI makes it easy to answer a hard email in ninety seconds. That is not always a good thing. If a message made you angry, the draft the model produces from your angry bullets will be a polished version of your anger. Sleep on it, or at least walk to the kitchen first.

Do not let it reply for you

Auto-replies drafted and sent by an assistant without your review are a bad idea for anything beyond "I am out of the office." The tools that offer this keep improving, but the failure is asymmetric: one wrong commitment to a client costs more than the minutes you saved on fifty routine replies.

A pre-send checklist

Thirty seconds, every time:

1. Is every fact, date, and number in this draft one I gave it, or one I have checked?
2. Does it commit me to anything I did not decide?
3. Does the first sentence say the thing that matters?
4. Would I say this out loud to this person?
5. Is there anything confidential in here that should not have gone into the tool?

If all five pass, send it.

Next steps

• Grab ready-made prompts from writing and email and customer communication.
• Set up a permanent voice assistant using Custom GPTs, Claude Projects, or Gemini Gems.
• Read AI privacy at work before pasting anything with client details.
• See how executive assistants and customer service reps use these techniques all day.
• Tool pages: Grammarly, Microsoft Copilot, and Gemini.
• Sharpen the prompts themselves with the prompt engineering guide.
• The course has a full writing and communication module with exercises.