Outbound becomes engineering when you build it as one measured system: a sending domain that has earned its reputation, a prospect base where every field has a source, and a gate that can refuse a send. Deliverability comes first, because a message that lands in spam is never read, however good its text is.
What outbound engineering is
Outbound engineering means building and running unsolicited B2B outreach as a measured system. I have run outbound since 2015, and volume was never what made it work. I think of it as a data pipeline with an email at the end.
The definition comes from a preprint of mine, listed on the research page. It names three properties that a system must have to qualify: provenance for every address, a record of each send that is written before its outcome arrives, and a gate that can refuse a send. Three more appear in every system I have built: isolation, volume that follows a reading, and sequences that follow behavior. The preprint treats those as design hypotheses, not as part of the definition.
The work has three layers: the sending infrastructure, the prospecting engine and the funnel that catches the demand. For a shorter answer, read What is outbound engineering?.
How email deliverability works for cold email
A cold email usually fails before anyone opens it. Mailbox providers such as Gmail and Outlook place each message according to the reputation of its sender, and a new domain has not earned any. That is why I begin with the domain and leave the message for later.
Two pieces of my own cover this part.
- email-intel is open source. Given a domain, it resolves who runs its mail against a 36-service signature database. It grades the anti-spoofing posture of the domain, which covers SPF, DKIM, DMARC and STARTTLS, with a letter. It ships as a library, a command line tool and an MCP server, so an agent can run the same checks.
- warming.email is the product on top. Its domains write to one another with real mail, a little more each day, answer one another and pull one another out of the spam folder. Readiness and health are exposed as APIs, so whether a domain can send yet is a fact you can query.
Two rules follow. The first is isolation. Prospecting gets its own subdomain with five to ten mailboxes, and the main domain of the company stays out of it. The second is that a reading sets the volume. Warm-up takes six to eight weeks. Through those weeks the system keeps measuring the reputation of each mailbox, and that measurement decides its daily cap: higher while the reputation holds, lower when it slips. In my systems a healthy mailbox settles at around sixty messages a day.
How a prospecting engine works
The engine has four stages: discover, enrich, score and outreach. It works on dossiers, not on rows in a spreadsheet. A dossier is one JSON document per company, versioned and written under a declared schema, and each field records where its value came from. A claim with no source stays out, which keeps an unsupported personalization out of every email.
- Discover casts wide, with agent-assisted web discovery and commercial data sources.
- Enrich fills the dossier with contacts, roles, site signals and fit data.
- Score applies a written rubric from 1 to 10. Band A, from 7 to 10, gets outreach. Band B, from 4 to 6, gets nurture. Below 4 is discarded.
- Outreach runs segmented sequences by email and LinkedIn. A compliance module sits in the path, for consent and opt-out rules.
Because the rubric is kept under version control, anyone can read later why a given company was contacted. The state of the pipeline lives in a queue file, and the dossiers live in versioned storage.
How campaigns and the handoff to sales work
Each campaign runs as an experiment. It opens with a short message in plain text, without links, that reads well on a phone. The next step depends on what the person did. Someone who never opened the message gets a nudge on day three and another on day eight. Someone who opened it without replying gets a different message. Someone who replied leaves the sequence.
Replies go to an agent that reads the inbox and puts each one in a class: interested, not now, wrong person, or a question. An interested reply is passed on at once. The sales team of the client sees it in Slack, in the CRM and on a shared calendar, with the context a salesperson needs to call before the thread goes cold.
Which addresses to email and when to hold a send
These are the rules I apply. Each one is a gate that can say no.
- A cold domain does not send campaigns. A warmed domain sends on a ramp, and the ramp is monitored.
- When the reputation reading of a mailbox slips, its daily cap comes down before damage is done.
- Send only to an address that the company published on its own site or that its mail server confirmed. An address inferred from a naming pattern is a guess, and guesses get no mail.
- When a change to targeting, volume or copy does not move the numbers, roll it back.
How to set up outbound step by step
- Write down the ideal customer profile: type and size of company, place, and the condition that makes a conversation worth having.
- Check the sending domain: who runs its mail and how its SPF, DKIM and DMARC posture grades.
- Create a separate subdomain with its own mailboxes and warm it up, with the reputation reading setting the daily cap.
- Build the prospect base while the domain warms: a dossier per company and a source for every field.
- Score each company against the written rubric. Only the top band moves to outreach.
- Write each campaign as an experiment: a short first message in plain text, then a next message that depends on what the person did.
- Measure deliveries, opens and replies in real Gmail and Outlook inboxes, by message, segment and sequence.
- Classify each reply and pass the interested ones to the sales team with their context.
What goes wrong with cold email
- The words come first. People start with the subject line, and the message is filtered before anyone reads it because the domain has no reputation.
- The main domain carries the campaign. One bad campaign then hurts the invoices and the support mail of the business.
- The list contains guessed addresses. A domain that took weeks to warm can lose its reputation in one week this way.
- Compliance is left out. Outbound that ignores consent and opt-out rules produces liability with every send.
A worked example: finding suppliers near new plants
I first built this engine inside one of my own products, and later I built it again for clients. The latest was an industrial real-estate developer in northern Mexico with a specific question: which suppliers are likely to need industrial space near a corridor where two global manufacturers were installing plants?
That question is the thesis, and the engine works from it. Discovery looks for companies that match. Each company gets a dossier, and each field keeps its source, down to the page that claims a company supplies one of those manufacturers. The rubric gives every company a score and a band. Only band A moves to outreach, with templates written for that thesis.
With the queue and the dossiers on record, the pipeline can be inspected after the fact. When a prospect converts, you can reconstruct which evidence produced the conversation.
The third layer catches the attention that outbound creates. For a B2B distributor of industrial supplies in the same region I built a quotation funnel that turns a quote request into a structured lead.
What the evidence shows and what it does not
- The system measures deliveries, opens and replies, so those are the results I can show.
- Signed contracts are a different matter. That step belongs to the sales team of each client, and I do not yet have figures from enough clients to report how replies become contracts.
- The preprint states a definition and seven propositions to test. It claims no empirical result, and it declares that I sell a managed outbound service under the same name.
- The weeks of warm-up and the daily volume in this guide describe systems I built. They are not a benchmark for another domain.
- I do not promise a reply rate or a delivery rate.