How to write better emails with AI

AI is useful for email when you treat it as an editor, not as autopilot. The best results come from giving the model real context — who the email is to, what they already know, what you want from them — and then using it to draft, revise, and tighten. The worst results come from typing "draft a follow-up email" and sending whatever appears.

This guide is for the email work you do every week: replies to prospects, internal asks, follow-ups after meetings, hard conversations, and rewriting your own first drafts into something that doesn't sound defensive or stilted.

A 5-part model for any email AI writes for you

The same five inputs make every email better. If you find AI-generated emails reading like LinkedIn cold spam, it's usually because one or more of these is missing.

Who is this to? Role, relationship to you, how formal they are, what they already know about you.
What do I want from them? A reply, a meeting, a decision, a forwarded intro, an action. Be specific.
What context do they already have? Last call, last email, the project they're on, the document they read.
What tone fits here? Direct but warm, formal, technical, casual, apologetic, firm. Tone fails are why "good" emails feel wrong.
What should happen next? The CTA. Reply by Friday, book 20 min, just acknowledge — make it unambiguous.

Give the model these five and it stops guessing.

A real example: bad input vs good input

Bad input:

Write a follow-up email to a prospect.

What you get: "I hope this email finds you well. I wanted to follow up on our last conversation regarding [project]. Please let me know if you have any questions. Looking forward to your response."

That email goes straight to trash. No specificity, no opener that matters, generic CTA.

Good input:

Write a follow-up to a prospect.

WHO: VP of Marketing at a 40-person B2B SaaS. We had a 30-min demo last Tuesday. She liked the product but said budget is tight Q1.

WHAT I WANT: A 15-min call to discuss whether a smaller-scope pilot would fit her Q1 budget.

CONTEXT SHE HAS: Saw the full demo, has the pricing page, knows our cheapest tier is $400/mo.

TONE: Direct, no pushing, acknowledge her constraint.

NEXT STEP: Reply yes/no on the pilot idea, or push to Q2.

Ban: "Hope this finds you well", "circling back", "I wanted to reach out", "looking forward". Cap at 90 words.

What you get: A short, specific email that opens with her constraint, proposes the smaller pilot, and leaves an easy out for Q2.

The bad input asked for an email. The good input asked for this email.

Use AI for revision, not just drafting

The biggest underused move with AI in email is rewriting your own draft. You wrote a paragraph that's almost right but sounds slightly off — defensive, too long, too soft, too aggressive. Hand it to the model with a clear instruction:

"Make this 30% shorter without losing any of the actual points."
"Remove the hedging language. I'm allowed to be direct here."
"This sounds defensive. Rewrite so it sounds like a competent colleague making an observation."
"Make this less salesy. It currently reads like a sequence email."
"Tighten the second paragraph — it has three ideas mashed together. Pull them apart or cut two."

Revision-mode is where AI is most useful. You already have judgment about what the email should say. The model is good at executing the rewrite cleanly.

What bad AI email looks like

You can spot AI-generated email from across a room. The shape is:

Opens with "I hope this email finds you well" or "I wanted to reach out"
Long sentences with too many subordinate clauses
Vague compliments ("impressive work at [company]")
Generic value statements ("I think there's an exciting opportunity")
"Looking forward to your thoughts" closer
Word count near 200 when 60 would have done it

The buyer's eye is trained on this pattern. The moment one of these surfaces, the email is marked as automated and discarded.

A ban-list to use in every email prompt

Tell the model not to use these. Adjust to your industry, but the universal list is solid:

"I hope this email finds you well"
"I wanted to reach out"
"Circling back" / "Just circling back"
"Touch base"
"Synergies", "leverage", "ecosystem", "best-in-class"
"Looking forward to your thoughts/response"
"Bandwidth" (used to mean availability)
"Hope you're doing well"

When you put this list in the prompt, the model produces tighter, more direct text. Almost every B2B email-prompt benefits from it.

Where AI helps with email

First drafts when you know what you want to say but not how.
Tone adjustment — making the same content read direct, warm, formal, or apologetic.
Length cuts — turning a 200-word paragraph into 80 words without losing what matters.
Hard emails — bad news, declines, pushbacks. Drafting these once removes the staring-at-screen tax.
Variant writing — three openers, pick one. Two CTAs, pick one.
Translation across formality levels — the email you'd send a peer vs the one you'd send their boss.

Where AI goes wrong with email

Relational context the model can't see. "How is the working relationship?" matters. If you don't tell it, it can't write the email correctly.
Cold spam patterns. Without explicit anti-patterns, models default to LinkedIn-DM voice.
Subject lines. Models write generic subjects. "Following up on our chat" is universal AI tell. Write subjects yourself or give the model strong examples.
Inside jokes and shared history. A line that should reference a thing you and the recipient both remember will get replaced with a generic compliment.
Tone you can't see. Sarcasm, dry humor, gentle pushback — these need explicit examples in the prompt, or the model defaults to flat.

When not to use AI for email

A few categories where AI is not the right tool:

Sensitive HR conversations. Performance issues, terminations, complaints. The tone has to be exactly right and the wording is legal-adjacent. Draft these yourself.
Conflict and difficult relationships. When the relationship is the point, the model can't see what you can see. AI tends to over-polish, which reads as cold.
Legal and contract-adjacent. Anything that creates obligations, commitments, or could be quoted back to you in a dispute. Write it; have it reviewed by someone qualified.
Apologies. A sincere apology is the place where AI's smoothing tendency does the most damage. People can tell.
Anything where being wrong about the recipient is expensive. AI doesn't know they got promoted last week, or that they're going through a divorce, or that they hate the word "circle back."

Common mistakes

Skipping context. "Write a follow-up email" without the five inputs above produces template-grade output.
Accepting the first draft. First drafts are 70% there. The second pass with a sharp revision instruction makes them shippable.
Forgetting the ban-list. Without it, AI defaults to the most generic email voice on earth.
Using AI for the email when the email isn't the problem. If you don't know what to ask for, AI can't write the email that asks. Figure out the request first.
One prompt for everything. "Write the email AND the subject AND a follow-up if no reply." Output is mediocre across all three. Do them separately, each with proper context.
Trusting AI on names, titles, and dates. If the model has to remember a specific fact about the recipient, paste it in. Don't let it guess.

The summary, plainly

AI doesn't write better email than you. It writes the version of your email that you would have written if you'd had another 20 minutes. The leverage is in revision, tone control, and getting unstuck on the hard ones — not in autopilot drafting.

Related:

/prompts/sales — cold outreach, follow-up sequences, objection handling, win-back emails
/prompts/communications — internal communications, status updates, hard conversations
Best AI prompts for sales reps — the deeper version for sales-specific email patterns
How to verify AI output before you trust it — when AI claims a fact about your recipient, check it