I built a self-feeding lead engine

So we just shipped something remarkable this week.

A cold-email engine that keeps a campaign topped up on its own. You set one number, a daily send target, and the system figures out how many fresh leads it needs, pulls exactly that many from the lead "warehouse," cleans them, writes a custom first line per lead, and loads them in paused. Ready to send. No one touching it.

But the part worth talking about is not the automation. It's the cost fix we found inside it.

Here's what we were doing before. We'd send a batch of leads to the AI model and ask it to personalize each one. Sounds fine. Except roughly 85 percent of every batch were consultants and services companies, leads that don't fit our clients' offerings. So we were paying the model to read those profiles, think about them, and then reject them. Over and over.

85 cents of every dollar spent on the model was just buying rejections.

The fix sounds obvious in hindsight. Filter in the database query before any model call even happens. Those service companies never make it to the model now. We only pay to personalize leads that already passed the fit test.

So the cost per qualified lead dropped significantly, and I mean the math is just cleaner across the board.

And then there was a second thing, a live-run lesson. We did dry runs for days, and everything looked fine. First real send, two silent bugs surfaced immediately. The engine was ignoring the send target I set in the portal and falling back to a default because I called the database driver the wrong way and the error got swallowed. And every upload was blocked by an assignment gate I never gave the campaign accounts to.

Neither showed up in any dry run.

The lesson I'm keeping from that: a dry run proves the math but not the plumbing. The first real send is its own test, right?

You have to build expecting that. So if you're using AI to personalize at scale, check what you're asking the model to process before it ever reads a word.

The filter belongs in the query, not the prompt.

talk soon

Roman

Tell me what you're really selling, who you're selling to, and what you've already tried. 7 questions. 4 minutes. I'll read your answers personally and reply with what I'd test first. Start here.