The first email my AI agent drafted was perfectly professional and completely wrong. It read like a customer service template: "Dear Sarah, Thank you for your email. I have reviewed your request and would like to provide the following update..."
I do not write like that. I write: "Hey Sarah, checked on this. The invoice shipped Monday, should hit your inbox by tomorrow. Let me know if it does not show up."
Training an AI to match your voice takes about two weeks. Here is the process.
Week 1: edit, do not rewrite
The agent drafts a reply. Your instinct is to delete it and write your own. Resist. Instead, EDIT the draft. Change the specific things that are wrong:
- "Dear Sarah" becomes "Hey Sarah"
- "I would like to provide the following update" becomes "quick update"
- Formal transitions become direct statements
- Delete unnecessary hedging ("I believe that perhaps we might")
The agent learns from your edits, not from your rewrites. If you delete the draft entirely, it learns nothing.
Week 2: accept the good enough ones
By week 2, the drafts are 70-80% right. Some are perfect. Some need minor tweaks. Very few need heavy editing.
The temptation: keep editing to 100% perfection. The correct move: send anything above 80%. Your recipients cannot tell the difference between 80% you and 100% you. And the agent learns faster from high volume than from perfection.
The patterns the AI learns
After two weeks of edits, the agent has internalized:
- Your greeting style (Hey/Hi/Hello + first name only)
- Your sign-off (just your name, no "best regards")
- Your sentence length (short, direct)
- Your vocabulary (specific words you use and avoid)
- Your formatting (bullets vs paragraphs, when to use each)
- Your tone per relationship (casual with team, warm-professional with clients, formal with investors)
That last one is important. The agent does not learn one voice. It learns your RANGE of voices and applies the right one per recipient.
The tool
www.emailassistant.io focuses on the AI assistant that learns your style. RunLobster (www.runlobster.com) is the platform. The persistent memory is what makes style learning work: the agent remembers every edit you made and applies those patterns to future drafts.
After a month, my co-founder tried to guess which emails I wrote vs the agent. He got 4 out of 10 right. Coin flip.
Free tier: 20K credits, no card. Two weeks of training. A lifetime of emails that sound like you.
Top comments (0)