TL;DR: do not automate sending, automate the blank page. A rep writing follow-ups by hand manages about 20 a day; reviewing and approving drafts they manage 80 with the same personalisation. Build: three to four days. The approval step is not a safety net you remove later, it is the thing that makes the output good.
The bottleneck in follow-up is not judgment, it is typing and the reluctance to start. Reps know they should follow up. They do not, because attempt three is genuinely hard to write and there is always something more urgent.
So leads that showed real intent go cold, deals stall, and revenue leaks quietly enough that nobody has to make a decision about it. Your CRM shows you the problem perfectly and changes nothing.
The pattern: draft, approve, send as the rep
A deal sits idle past a threshold you choose. The system pulls the context that matters - company, contact, last note, amount, stage, days quiet - and a model drafts a short follow-up referencing what was actually discussed. The draft goes to the rep with an approve action. They read it, edit if needed, approve. It sends from their own mailbox and the CRM logs the activity.
Ten seconds per email instead of ten minutes, and the person whose name is on it still read it before it went.
Why the approval step stays
Teams always ask why not send automatically, and the answer is not fear of a bad email. Three reasons, in order of importance.
The rep owns it. Someone who clicked send behaves differently from someone the system emailed on behalf of, and that shows up in the call that follows.
The rep catches what the CRM does not know. Half of what matters about a deal was said on a call and never written down, and the model cannot see it.
Every edit is a free signal. Watch which drafts get changed and you learn exactly what the prompt is getting wrong, without running a survey.
It is also the thing that keeps this on the right side of the line between follow-up and bulk sending, which matters legally as well as commercially.
Why drafted beats templated
Templates get ignored because they read like templates, and buyers have seen thousands. A draft written against the actual deal references the objection raised last time and proposes a next step that fits the stage. Same effort from the rep, different response rate.
What goes in the instruction, specifically
Most of the quality lives here, not in the model choice.
A hard length limit. Four sentences. Long follow-ups do not get read.
Reference exactly one specific thing from the last interaction, not a summary of the relationship.
Propose one concrete next step with a time, not a request to reconnect.
Forbid inventing anything not in the supplied context. This is the rule that prevents a confident reference to a conversation that never happened.
Forbid the standard filler: circling back, touching base, hope this finds you well.
If there is nothing new to say, output nothing. A system willing to produce no draft is far more useful than one that always produces something.
That last one is the highest-leverage line in the prompt. The fourth follow-up with no new information should not be sent at all, and a model instructed to always write will always write.
Where the threshold should sit
Seven days of no activity is the common default and it is a reasonable place to start. Worth tuning by stage: a deal that went quiet right after a proposal wants a different interval from one at first contact. What matters more than the number is that the clock runs on real activity - email, call, note - rather than on the record being touched, or reps will learn to touch records.
What you need
Nothing exotic. The build is three to four days including testing.
A CRM with API access: HubSpot, Pipedrive, Salesforce
An automation platform, Make or n8n, as the orchestration layer
A model API key
Mailbox connection via OAuth, never a stored password
Wherever your team already works, for the approval step
The limits, which are not optional
This is follow-up to people already in your pipeline, which is a different legal position from cold outreach. Three things still apply and they differ by market.
An unsubscribe that works and a suppression list that every channel respects, permanently. Build it on day one, not after the first complaint.
In the EU and UK, a documented lawful basis. Existing-customer follow-up is a different case from a lead who filled a form once.
In Australia, accurate sender identification and a functional opt-out under the Spam Act, every time.
The approval step helps here too: a human seeing each message is a meaningful difference from bulk sending, both in how it is received and in how it is characterised.
What the numbers looked like
On the B2B system in our case study, 60-80 drafts a week. Reply rate moved from around 12% on manual, inconsistent follow-up to around 28% on consistent drafted follow-up. Deals that would have stalled two or three weeks were getting responses inside five days.
Read that improvement carefully, because it is easy to over-attribute. The larger share of it is consistency, not wording: follow-ups that previously did not happen at all now happen. Any system that reliably sends the second and third touch will beat one that depends on someone remembering, whoever writes the words.
Want to see the shape of this on your CRM? The audit at 2pizza.team/audit takes two minutes and no call.
Originally published at 2pizza.team. We build AI and automation systems for small teams - fixed price, two to six weeks. See the work.
Top comments (0)