The most common outbound failure we get called in to fix is not bad copy or a bad list. It is a team that scaled volume on infrastructure built for a tenth of it, and is now watching reply rates fall for reasons nobody can explain.
The fix is architectural and needs to happen before the volume, not after.
The arithmetic
A safe cold-sending ceiling is roughly thirty to fifty messages per mailbox per day. That number has not moved much in years and is not a limit your sending tool imposes — it is where recipient-side pattern detection starts noticing.
So two thousand sends a day is forty to sixty mailboxes. There is no configuration that makes five mailboxes do that job. Teams that try discover it as a cliff rather than a slope: everything looks fine, then a domain is effectively dead.
Domain architecture
Never the primary. Register variants — the hyphenated form, the .co, "get" or "try" prefixes — and treat them as consumable infrastructure. Three to five mailboxes per domain; more concentrates risk, fewer wastes warm-up effort.
Authenticate each one independently: SPF, DKIM, and DMARC actually enforcing. A domain with DMARC at p=none is announcing that it does not police itself.
Age them before use. A domain registered yesterday sending cold mail today is the strongest negative signal in the system. Two to four weeks of age plus a warm-up ramp is the minimum, and it is why the estate has to be planned a month ahead of the campaign that needs it.
Treat sending domains as consumable infrastructure. If one gets damaged, you retire it and nothing else notices.
Warm-up is a ramp, not a switch
Start at a handful of sends a day and build over four to six weeks. Keep some baseline sending going permanently — reputation decays with inactivity, and a mailbox that sat idle for two months is not the mailbox you warmed.
On reciprocal warm-up networks: they are increasingly detectable and the direction of travel is against them. Use them as a supplement to genuine engagement at most, and never as the entire warm-up.
Rotation that does not look mechanical
Distribute sends across the estate, but distribute them like a human would. Randomise intervals rather than sending on a fixed clock. Keep each thread on one mailbox — switching senders mid-conversation is both a detection signal and confusing for the recipient. Send inside the recipient's business hours, not yours.
Monitor per mailbox, not in aggregate
The aggregate number is what hides the problem. A campaign averaging a healthy reply rate can contain six mailboxes at zero, and by the time the average moves those six are gone.
Track reply rate and bounce rate per mailbox weekly. Pull anything that halves. Seed-test placement per provider monthly. This is fifteen minutes of work that regularly saves a quarter.
What it costs
A fifty-mailbox estate is a real line item — domains, mailboxes, warm-up, and the tooling to manage it. Teams balk at it and then spend more than that on data for a list they cannot safely mail.
The order of operations is the whole lesson: infrastructure, then list, then copy. Reversed, which is how most programmes are built, the infrastructure becomes an emergency.