If you write a lot of business emails, you've tried using AI to draft them. The output is usually fine — grammatically correct, structurally reasonable, totally forgettable. The recipient's response rate doesn't change. The hours saved drafting come at the cost of an extra reply cycle when the recipient asks clarifying questions you would have anticipated.
Better use of AI for email starts with knowing which parts AI is good at and which it ruins.
What AI does well for email
The unblocking moment. Staring at a blank reply window? AI gives you something to react to. The first draft doesn't have to be the final version; it just has to break the inertia.
Tone calibration. "This email reads angry. Rewrite it to be firm but professional." AI is excellent at this surgical tone work. Faster and less embarrassing than asking a colleague.
Long-thread summarization. "Here's a 12-message thread; summarize what's been agreed, what's pending, and what each party owes." AI does this well; the alternative is reading the whole thread again.
Drafting boilerplate. Confirming meetings, sending receipts, following up on requests — repeatable patterns you don't need to think about.
Translation for cross-language work. If you work across English / Chinese / Japanese, AI translates better than DeepL on email-style writing now. Edit for tone but the translation core is reliable.
What AI ruins
The opening line. Generic openers ("Hope this email finds you well") are red flags. AI's defaults make you look like you didn't read the previous message.
Specific references. "Per our conversation last Tuesday at the offsite, when you mentioned your team is rolling out the new evaluation framework Q2..." — these moments demonstrate you remember and care. AI invents specific references or stays vague enough that you sound generic.
Brevity. AI defaults to 3-4 paragraphs when 2 sentences would close the loop. Most professional inboxes are overwhelmed; people respond faster to shorter messages.
Personal voice in long-term relationships. People you email regularly recognize your patterns. AI sounds like you-but-flattened to anyone who reads more than 2 of your emails.
Difficult conversations. Performance feedback, declining favors, addressing conflict — emails that require emotional intelligence. AI either softens too much (so the message doesn't land) or sticks to corporate-blandness (so you sound like HR).
A useful framework
For each email, ask:
- Does the recipient know me? With strangers, AI defaults are often fine. With colleagues, AI defaults sound off.
- What's at stake? Low stakes (confirming a meeting): AI is fine. High stakes (asking for a raise, declining a partnership): write yourself.
- Is this a relationship I want to deepen? Cold outreach: AI fine. Building rapport: write yourself.
- Does this need precise language? Technical specs, legal notices: AI assist + careful review. Emotional nuance: write yourself.
The pattern: AI for transactional, you for relational.
Better prompts for email
Draft a [length] response to this email. The recipient is
[relationship to me]. I want to convey [main message]. My tone
should be [warm/firm/neutral]. Do not include "Hope this finds
you well" or similar. Reference these specific details from
their message: [list].
Here's the original email: [paste]
Here's relevant context they don't have: [paste]
The specificity matters. Each constraint narrows AI away from default-bland output.
Common AI email failures
- "Per my last email" — passive-aggressive but AI uses it without sensing the connotation.
- "As discussed" — when nothing was discussed.
- "I wanted to reach out" — useless preamble; just say what you want.
- "I hope this email finds you well" — universal cringe.
- Bullet lists when prose would be human. AI loves bullets; sometimes a paragraph is more direct.
- Closing with "Looking forward to hearing your thoughts" when no thoughts were requested.
Build a personal banned-words list for your email AI prompts.
Long emails: when AI helps and hurts
For very long emails (project updates, post-meeting recaps, executive briefings), AI's structuring is genuinely useful:
- Break into clear sections
- Highlight decisions and action items
- Move questions to the top so people answer them
- Cut filler sentences
But: long emails often shouldn't be long. "Could this be a Slack message? A 2-paragraph email? A meeting that's 10 minutes shorter?" Test before drafting; AI will always produce email of whatever length you ask, even when no email should exist.
Privacy and compliance
Don't paste these into ChatGPT or any non-privacy-tier AI:
- Emails containing PII (other people's personal information)
- Confidential business strategy emails
- Anything subject to attorney-client privilege
- HR matters: performance reviews, hiring decisions, terminations
- Health information (HIPAA-relevant)
Use enterprise tiers (ChatGPT Team / Enterprise, Claude for Work, Microsoft Copilot for Business) where data isn't used for training. Or run a local model. Or write the sensitive part yourself.
When NOT to use AI for email at all
- Performance feedback (positive or negative) for someone who reports to you
- Difficult conversations with people you have ongoing relationships with
- Apologies that should feel sincere
- Resignations, layoffs, exits
- Any conversation where the words specifically chosen carry meaning beyond the content
For these, write yourself, even if it takes 20 minutes more. The email itself is the relationship management; outsourcing it to AI is outsourcing the relationship.
Decision tree
- Quick transactional email: AI draft + 30-second edit
- Cold outreach: AI draft + personal customization
- Reply to ongoing relationship: write yourself, AI for tone check
- Performance / difficult conversation: write yourself, no AI
- Long status update: AI structures, you fill specific content
Next steps
- Build a personal style guide for your AI email prompts (banned words, preferred phrasings)
- Try drafting your next 5 emails three ways: AI-only, AI-then-edit, you-only. Notice which got better replies.
- Read about email habits like the 27/9/3 rule (sentence count for replies)
- Use AI for the parts that drain you; never for the parts that distinguish you