AI cold email writing, trained on 4M sends | Empra Outbound
Emails written on science, not taste.
We spent three years and four million emails finding out what actually gets answered. The model learned it. Now it writes with it, for one person at a time.
The corpus
Four million sends taught it.
Three years of our own mail. Not a public dataset, not scraped. What went in was every email we sent and what came back; what comes out is the four decisions the engine makes about every email it writes.
- 4,000,000 EMAILS SENT BY US, OVER THREE YEARS
- 650,000 LINKEDIN MESSAGES
THE ENGINE
- SUBJECT
- CTA
- TIMING
- BODY
What it learned
The science of a cold email.
Every send came back as data. What got opened, what got answered, what got deleted. The model turned that into rules it writes with.
- Subject lines
Which openers get opened, by which persona, at which seniority. - Length
How many words earn a reply before attention runs out. - Call to action
Soft ask or hard ask, meeting or question, per persona. - Structure
Where the personalisation sits, where the point lands. - Send window
Day and hour, per mailbox, per timezone. - Cadence
The gap between touches, and how much runway a variant gets.
It has never seen your market, so the angles still get tested rather than assumed. And it does not decide whether your offer is any good. Nothing does that but your offer.
Per lead
Six or more variables, then a discard list.
The engine finds far more than it uses. A detail that is true but irrelevant reads as surveillance, so it gets thrown away.
Written into the email
- Second req open since the 14th
- Job spec asks for outbound
- One AE on the team
- Site sells to ops leads
- Round closed in March
- Second persona, derived
Thrown away
- Alma mater of the VP Sales
- Podcast appearance, two years old
- Office move
- Investor names from the round
- Personal marathon time
True, and useless. Most agencies never show you the second column.
How an email gets written
Six passes, every time.
The same sequence runs for every single lead. It does not get skipped when the list gets big.
- Read the company
Research site, product, hiring, round, stack. The agent goes wide before it goes narrow. - Keep six things
Filter the few facts that touch your offer. Everything true but irrelevant gets discarded. - Pick the angle
Choose which of your tested angles this particular company is most likely to answer. - Write to the person
Draft one email, built from what was found. Never a template with a merge field in it. - Apply the science
Subject, length, call to action and send window come from four million sent emails, not from taste. - A person reads it
Nothing sends until somebody has looked at it. That gate never comes out.
Try the rules yourself
The same rule-set, free, on your email.
The model’s mechanical rules run in the open: paste your cold email and see what it would keep and what it would cut.