Case studies / The free lead engine
I wired four cheap tools into one pipeline that scrapes leads for nothing, throws most of them away on purpose, and emails the rest. Then I handed the whole thing to a bot so I would stop touching it. Here is how it fits together.
Cold outbound is mostly janitorial work. You find the people, you check that their emails are real, you decide who is actually worth contacting, you write the first line, and then you babysit the sender so it does not torch your domain. None of those steps are hard. They are just slow, and the usual fix is to pay a different tool for each one and a person to run the seams between them.
I wanted to see how much of that I could move into software end to end. Scrape, verify, qualify, send, and manage. The constraint I gave myself was simple. The acquisition cost per lead should round to zero, and the judgment work should be done by a model, not by me.
Five stages, one direction. Each stage hands a smaller, cleaner list to the next. The first four build and send the list. The fifth one runs the whole thing on a schedule so it keeps going without me.
The front of the pipeline is LeadScrape. You point it at a niche and a geography, say independent insurance agencies in a handful of states, and it pulls a list of businesses with names, websites, role titles, and contact emails. The acquisition cost per record is effectively nothing.
What comes out is raw and messy. Plenty of role addresses, plenty of dead inboxes, plenty of companies that are nothing like who I want to reach. That is fine. The whole point of the next three stages is to be ruthless about throwing this list away.
Every raw email goes through Reoon before anything else happens to it. Reoon checks whether the address actually accepts mail and flags the ones that do not, the catch-all domains, and the disposable junk. Anything that is not cleanly deliverable gets dropped here.
This is the step that makes the pipeline worth building. A verified list is still just a pile of contacts. It does not know who fits and who does not, and it has no idea what to say to any of them. That judgment is the expensive part of outbound, and it is the part I wanted a model to own.
Each verified row goes to Claude with a short rubric that describes the ideal customer. Claude reads what is known about the lead, the company, the role, a snippet from their website, and does two things at once. It scores how well the lead fits the rubric, and for the ones that pass, it writes a specific opening line based on what it just read. The leads that do not fit are dropped with a reason.
This is where the list collapses, and that is the goal. A verified address only tells you a message will arrive. The rubric tells you whether it should. By the time Claude is done, a list of thousands is a list of hundreds, and every survivor already has a first line written for it.
Here is the attrition across one representative run. The shape matters more than the exact numbers. Most of what you scrape is supposed to disappear, and the biggest single cut is the qualification step, not the verifier.
The survivors, now verified, qualified, and each carrying their own opening line, get uploaded into Instantly. The sending is spread across a pool of warmed inboxes so no single mailbox carries too much volume, and each lead enters a short sequence rather than a one-shot blast. Because the personalization was written upstream by Claude, every first email is specific to the person receiving it without anyone sitting there typing.
The pipeline so far builds and sends a list. The last piece is what keeps it alive. A second agent runs on a schedule and does the operator job that a person would otherwise do by hand. It reads the campaign stats, decides what needs to change, makes the change, and routes anything that looks like a real reply to me.
Day to day it is reading numbers and making small corrections. Pause an inbox that is bouncing, rotate sending volume, push a lead into a follow-up step. Once a week it does the bigger job of rewriting the copy based on what actually got replies. The leverage is that the boring, daily, easy-to-skip work gets done every single day, because nobody has to remember to do it.
The usual outbound stack pays for the same five jobs with four subscriptions and a person. This one moves the judgment into a model and the operations into a scheduled agent, and the only real recurring cost is the per-check verification and the model calls.
The real takeaway is not the specific tools. It is that the two expensive parts of outbound, the judgment about who is worth contacting and the discipline to run the campaigns every day, are both things software can now do well. Once you accept that, the rest is plumbing.
If you are drowning in the manual parts of outbound, this is the kind of system I build and hand over. Tell me your niche and what a good lead looks like, and I will tell you honestly whether a pipeline like this fits.
Tell me about your outboundInfrastructure, deliverability, lists, copy, and sequences.
When splitting work across agents actually pays off.
How a bot ends up running something every morning.
Advisory and build engagements on systems like this one.