What we learned shipping the first million AI follow-ups

Six things you can only learn at scale

We've now sent more than a million AI-generated follow-ups across 142 SaaS funnels. Here are the findings that surprised us most.

1. Subject lines barely matter

Inside the first 90 seconds of relationship, who the email looks like it's from matters more than what it says. Get the sender identity right and subject A/B testing is largely noise.

2. Reply rate is the only real KPI

Opens are vanity. Clicks are an okay proxy. Replies are the only signal that survived correlation against revenue across every customer cohort we ran.

3. The best emails are short and ask a question

Median high-converting email length: 41 words. Median low-converting email length: 187. Every word past 60 cost reply rate.

4. Founders write better than copywriters

When we let founders leave a voice note describing the user's situation, the resulting AI email outperformed the same model briefed by a senior copywriter — by 22%.

5. Auto-send beats draft-mode in week three

Teams that started in draft-mode and graduated to auto-send within three weeks saw 2.4× more conversions than teams that stayed manual. The model just makes fewer bad calls than a tired human.

6. Don't automate the hard reply

When a user replies with a substantive question or objection, the AI should hand off to a human. Every team that auto-replied to objections lost trust, fast.

The lesson: automate the boring 80%, protect the human 20%.

What's next

We're rolling these findings into the default VibeFollow loop — short, founder-voiced, reply-aware, and quick to escalate. If you want the practical playbook, hit reply.