Give an agent a prospect list and a sequence template, and it will happily send. That's the problem. "Can send" and "should send" are different questions, and most outreach automation never asks the second one.
The fix isn't a smarter sequence. It's a gate the agent can't route around: no research artifact, no send — enforced in the pipeline step, not suggested in a prompt.
Why the gate has to be structural
The failure mode isn't laziness, it's defaults. An agent asked to "personalize and send" will reach for the fastest personalization it can produce — {{first_name}}, a line about the company's industry pulled from the domain, maybe a guess at their role. None of that is research. All of it reads as research to a prospect for about one sentence, then reads as a mail merge.
Telling the agent "do real research first" in a system prompt doesn't survive contact with a 500-row list and a deadline. The instruction competes with every other instruction in context, and "send" is a satisfying terminal action an agent is biased toward reaching. The only way personalization-before-send survives is if the pipeline literally has no send step that accepts a lead without an attached dossier.
What the gate actually checks
A dossier, for this purpose, isn't a paragraph of flattery — it's three to five concrete facts an agent had to go look up, not infer from the domain string:
- A specific signal tied to why now — a hire, a launch, a funding round, a tool change visible from job postings or changelogs
- Something true about the company that a template couldn't have guessed
- The actual person's role and a plausible reason the offer intersects their job this quarter
If the enrichment step can't produce that — the company's too small, too private, or nothing distinguishes them from the last fifty rows — the gate's correct behavior is to drop the lead, not to downgrade to generic. A sequencer that lowers its bar under volume pressure is worse than one that sends fewer, better emails.
The second gate: deliverability, checked before volume, not after it tanks
The dossier gate stops bad personalization. A second, separate gate stops the thing that kills the whole domain: ramping send volume before the domain has reputation to spend.
The checklist that matters, roughly in order:
- Separate sending domain from your primary — a burned subdomain is recoverable, a burned root domain costs you email entirely, including transactional mail.
- Warmup schedule before any cold volume — days of low-volume, high-engagement sending (replies, opens) before the sequence is allowed to touch its target list size.
- Volume caps that ramp, not jump — the agent enforces a ceiling per day that increases on a schedule, not on your impatience.
- A spam-trigger pass on the copy itself — the obvious stuff (ALL CAPS subject lines, link-only bodies) still shows up in agent-generated copy if nothing checks for it.
Here's the part worth stealing even if you build nothing else from this: a scheduled, unattended version of this pipeline should refuse to schedule itself until warmup is complete. Not warn — refuse. An agent that can talk itself past "maybe just this once" during warmup is an agent that will talk itself past it on the cap, too.
Reply triage is the other half of "don't waste a human's time"
The enrichment gate protects the send. A classifier on the way back in protects your attention. Every reply gets bucketed — interested, objection, or decline — before a human ever opens the inbox, so the only threads that reach you are the ones where a reply from you is the next useful action. An objection-handling library (price pushback, timing, "not the right contact") answers the second-most-common case without a human draft at all; a human only drafts the genuinely novel objection.
The combination is the actual point: an agent can safely own enrichment, drafting, and triage end-to-end, because none of those steps commit you to anything external. The one step that does — pressing send to a real stranger — is the one step kept out of the agent's hands by design, not by discipline you have to remember to apply.
Building this yourself
None of this needs special tooling — a skill that refuses to pass a lead downstream without a dossier field populated, a cron job that checks "days since warmup started" before releasing volume, and a classifier prompt with three labels will get you most of the way. We packaged a version of it — enrichment agents, 12 sequence templates, the warmup schedule, and the reply classifier — as Outreach Machine because building and testing the gate logic took longer than writing the sequences ever did. But the gate is the idea worth keeping regardless of what sends the email.
Top comments (0)