A few weeks ago I tried to explain to some friends what I've been building, and it came out as a list. A warming service. A browser tool that turns web pages into data. A subdomain. A queue of emails. An agent that watches an inbox. A Slack channel. A shared calendar. Somewhere around item four, one of them replied "lol", and honestly, fair. Told that way it sounds like a pile of tools held together with enthusiasm.
It's not a pile. It's one machine. And the shortest honest name I've found for the work of building it is outbound engineering.
The words are the last thing that matters
Everyone who tries cold email starts with the words. What should the subject line say? How long should the first paragraph be? That's the wrong end of the problem, because most cold email fails before a human ever sees it. Gmail and Outlook decide where a message lands based on the reputation of whoever sent it, and a fresh domain has no reputation at all. You can write the best email of your life and it'll sit in a spam folder next to the discount pharmacies.
So the work starts underneath the message. For me it started as a personal annoyance: I wanted an email from a new domain to reach a real person's inbox, and I found out that nothing about that was easy. So I built a setup where a pool of domains send real mail to each other on a slowly rising schedule, reply to each other, and rescue each other from spam, while I keep track of how every one of them is being treated. That became warming.email. It exists because you have to earn the inbox before you get to use it.
Warm up a subdomain, not the company
Rule two is isolation. Outbound never runs from a company's main domain. It runs from a separate subdomain with its own five to ten mailboxes, so if something goes sideways, the invoices and the support replies of the actual business keep landing where they always did.
Those mailboxes warm up for six to eight weeks. The system isn't just waiting during that time. It measures reputation constantly and uses that reading to decide how much each account is allowed to send that day. The number isn't a guess. When the reading says a mailbox can take more, the cap goes up. When it dips, the cap comes down before any damage is done. A healthy mailbox settles around sixty messages a day, so a full set lands somewhere between three hundred and six hundred a day, all of it arriving where it should.
Nobody loves this part. Six to eight weeks is a long time to wait before the first real email goes out. It's also the entire difference between a campaign that reaches people and a campaign that reaches filters.
A list is not a database
While the domain warms, we build the prospect base. The starting point is the client's ideal customer profile: what kind of company, what size, where, and what has to be true about them for a conversation to be worth anyone's time. From there the system finds companies, fills in public business information about them, and scores each one from one to ten against a written rubric. Only the top band gets outreach. The rubric is a document, so "why did we write to these people?" always has an answer.
Every field remembers where it came from. That matters most for the email address itself, because there are really three kinds. One published on the company's own site. One verified by asking the mail server whether it exists. And one guessed from a naming pattern, which is a guess wearing a tie. We only ever send to the first two. An inferred email is not an email. Sending to guesses is how a warmed domain loses its reputation in a week.
The client watches all of this fill up in a shared CRM as it happens, instead of getting a spreadsheet at the end.
Every campaign is an experiment
Once the infrastructure is ready, the campaigns start, and each one runs as an experiment, not a broadcast. The first message is plain text, no links, short enough to read on a phone. What happens next depends on what the person did. Never opened it? One nudge on day three, another on day eight. Opened it but didn't reply? A different message. Replied? Out of the sequence immediately, because the worst thing a system can do is send a scripted follow-up to someone who already answered.
Everything gets measured per message, per segment, and per sequence: deliveries, opens, replies, and what people actually did. The numbers come from real Gmail and Outlook, the same inboxes the prospects use, not from a sandbox that behaves nicely. Targeting, volume, and copy change based on those numbers, and a change that doesn't move them gets rolled back.
The handoff is where the value shows up
A reply is a signal, and signals go cold fast. An agent watches the inbox and sorts each reply: interested, not now, wrong person, has a question. For an interested one it can send over the information they asked for, request the details the sales team needs, or offer a link to book a meeting. In the same moment, the client's team gets the lead in Slack on their phones, in the CRM, and on a shared calendar, with enough context to call within sixty seconds, while the person is still looking at the thread.
That's the actual product. Not the emails. The moment a real conversation lands in front of a real salesperson while it's still warm.
Why "engineering" and not "marketing"
Line the pieces up and they read like a pipeline: profile, data, sending infrastructure, outreach, behavior, qualification, handoff. Every stage has a measurement, a feedback loop, and a gate that can say no. Reputation controls volume. Provenance controls who gets written to. Behavior controls the next message. That's the same discipline I use on evaluation harnesses for AI agents: an email that can't be traced back to evidence doesn't get sent, for the same reason a release that can't pass its gate doesn't ship. If you want the three infrastructure layers in more detail, I wrote them up as a case study.
From the client's side, none of the components exist. warming.email, the browser tool, the agent, the Slack bot: they're parts inside one acquisition machine. That's what makes the offer worth more than "I'll send five hundred emails a day for you," which is a sentence nobody should be proud of.
One honest caveat. I can show deliveries, opens, and replies with confidence, because the machine measures them. What I can't show yet, across enough clients, is how all that volume turns into signed contracts, because the sales team owns that last step and those numbers are still coming in. The frame is settled. The evidence at the far end isn't.
So here's the whole idea in one line. Reaching people is an infrastructure problem, and infrastructure is something you can build, measure, and improve. The machine takes a description of your ideal customer and turns it into conversations with real ones. Everything in between is engineering.