How to get AI email drafts that sound like you

You asked your AI email assistant to draft a reply. What came back was technically correct, professionally worded, and sounded nothing like you. The opening was "I hope this email finds you well." You would never write that — certainly not to this person.

The AI didn't fail. It just didn't know who you were writing to, or who you are when you write to them. That's a setup problem, not an AI problem.

Why AI email drafts sound generic

Without specific signals about you, AI defaults to professional person — correct but indistinct. But here's the subtler problem: even tools that do learn your general style apply it uniformly to every recipient. That's wrong. You write differently to your co-founder than to a new enterprise lead. Differently to your lawyer than to a longtime customer. Tone, length, what you assume they already know — all of it shifts by person.

A draft that sounds like you but doesn't sound like you talking to this specific person is still wrong. That's the gap most AI email tools never close.

What AI actually needs to write like you

1. Your sent mail history. Your past emails contain vocabulary, sentence length, opener patterns, sign-off habits, and formality level — hundreds of micro-signals you've never consciously articulated. Grant access to at least a year of sent mail. If your tool doesn't use sent history, that's why the drafts feel generic.

2. Recipient-specific context. This is where most tools fall short and where the real gap lives. You don't write the same way to your co-founder (clipped, shorthand, no pleasantries) as you do to an investor you just met (warmer, more structured, headline number first). With a customer escalation you're careful and never leave a question unanswered. None of this is consciously decided — it's accumulated over every exchange with that person.

A genuinely good draft sounds like you with this person — built from your specific history with that contact. When it works, the draft for your ops lead is terse and action-oriented. The draft for a first-time enterprise prospect has the right level of warmth and a clear next step. You open it and the edit is minimal — not because the AI got lucky, but because it learned the same way you did.

3. Full thread context. A reply is part of a thread with history and open questions. Good AI drafting tools read the whole thread and address everything that's been asked — not just the last message.

4. Learning from your edits. When you edit a draft before sending, that feedback should update the model. If your tool doesn't learn from edits, you'll be making the same corrections in month six as you were in week one.

How to set it up correctly

Week 1: Grant full sent history access (a year or more). Connect contacts so the AI knows which are investors, customers, and colleagues. Set coarse preferences — formality level, default length, sign-off style.

Weeks 1–2: Edit every draft, then send your edited version. You're training the model. If you keep deleting the opening greeting or cutting the final paragraph — those are patterns the AI should learn and stop repeating.

Week 3+: Drafts should now be close — under a minute of editing rather than a full rewrite. Use it for harder emails: cold outreach, sensitive customer communications, investor updates. The model is now strong enough to handle specific instruction well.

The number that matters: acceptance rate

Day-one acceptance rate for most tools is 20–30%. For tools with a compounding memory model, this climbs to 70–80% within 30 days. For tools without one, it stays at the day-one baseline. You're making the same edits in month six.

The single most important question to ask of any AI email tool: does it get better the longer I use it?

Common mistakes

Using it for high-stakes drafts too early. Start with routine, high-volume emails — meeting responses, follow-ups, confirmations. Get the model calibrated before using it for anything important.

Treating tone settings as the whole solution. "Casual" or "professional" is a coarse filter. Your sent history and learning loop are doing the actual heavy lifting.

Quitting before the ramp period ends. AI email drafting needs 2–4 weeks to calibrate. Three days is evaluating the cold start. Give it the ramp.

What to look for

— Does it analyse at least a year of sent history?
— Does it build per-recipient models, not just a global style?
— Does it learn from your edits?
— Does draft quality improve measurably at 30 days?

Faraday's AI drafting is built around exactly this: a persistent model of your voice, then per-recipient context on top — so every draft knows not just how you write, but how you write to this person. Most users are at 85%+ acceptance rates by day 15. By week three, it just sounds like them.