Nebo · tutorial 03 · win customers
Run outbound overnight, review it with coffee
A scored account list with drafted outreach waiting each morning, and not one message sent without you.
Outbound fails for two reasons: nobody has time to research properly, and generic messages get ignored. An employee working overnight fixes the first, which fixes the second. What it must never fix is your judgment about who to actually contact.
Steps
Write down your ideal customer profile first
Do this before hiring anyone. If you cannot describe a good-fit account in five bullets, the employee will score against a definition it invented, and you will spend a week arguing with its output instead of reading it.
Include disqualifiers, not just qualifiers. "Under 10 staff" and "already using a competitor we cannot displace" save more time than any positive signal.
Hire the SDR and connect the CRM
Install sdr into your sales department. Connect your CRM
and the mailbox you send from. Give it read access to the CRM before you give it
write access, and leave write off entirely for the first week.
Give it your real writing, not a description of it
Attach five outbound emails you actually sent that got replies. Not your best five; your typical five. It will imitate what you give it, so giving it your polished work produces messages you would not have the energy to write again.
Run it once, on a small list, by hand
Read all 25. You are checking two things: whether the scoring matches how you would have scored, and whether you would put your name on the drafts. Correct the scoring in the persona and the voice in the examples.
Split research from writing
Once it works, make the research step deterministic where you can. Pulling the
account list, deduplicating against active sequences, and checking suppression
lists are data operations, and data operations belong in command nodes rather than model activities. Let the model do the scoring judgment and the
prose, and hand it the list as a file it reads.
A model asked to both fetch and reason will occasionally invent a row. A command node's output is its stdout, byte for byte, so an account that is not in your CRM cannot appear in the list. On a 25-account run you would catch it; on a 300-account run you would not.
Schedule it overnight
Two in the morning, Tuesday through Saturday, so a list is waiting on the mornings you actually work. Set the employee's timezone input, or the cron will run in a timezone you did not intend.
Set the gate
Sending and sequence enrolment stay at approval permanently. Not
for a fortnight while you build trust: permanently. A routine that sends is a
routine that can apologise on your behalf to three hundred people before you
wake up, and outbound damages a domain reputation you cannot easily rebuild.
Reviewing drafts is a five minute job, and it is the five minutes that make the rest safe.
Verify
- Every account it scored highly is one you would have called
- No contact appears who has opted out or is already in a sequence
- The drafts reference something specific and true about each account
- You would sign your name to all of them without editing
- Nothing was sent
When it goes wrong
The drafts are all subtly the same
It found one angle and applied it everywhere. Usually means the research step returned thin material, so the writing step had nothing specific to work with. Fix the research, not the writing.
It scores everything as a good fit
Your profile has qualifiers but no disqualifiers. Add explicit exclusions and ask it to state the disqualifying reason when it skips.
It contacted someone
Check Controls immediately: outbound contact should be approval. If it is,
check whether the send happened through a different tool than the one you gated, for
example a CRM sequence rather than the mail plugin.