There is a version of "AI for sales" that is everywhere right now. You point a tool at the internet, it hands back a few thousand plausible-looking companies, and you start emailing. It feels productive. It is mostly noise. Half the accounts don't fit, a good number are your own competitors, and not one of the rows is backed by anything you can actually check.
We wanted the opposite of that. Not a bigger list, a list we could trust, where every company on it earned its place. So before we wrote a single line of automation, we wrote a rulebook.
The rulebook is the product
Finding companies is easy. The hard part, the part that separates a useful pipeline from a spam cannon, is rejection: deciding, with discipline, which companies don't belong and exactly why.
Ours started as one plain document: the standing conditions a company has to pass before it enters our list. A few of them:
- End-users only. Not consultancies, not agencies, not vendors who sell on the platform we target.
- Proof of current use, not a case study from three years ago. The strongest evidence we accept is a live posting on the company's own careers page for a role that names the tool. You don't hire a Salesforce admin for a CRM you don't run.
- A verbatim quote and a real source URL for every claim. No paraphrasing, no "trust me."
- No ecosystem vendors, no direct competitors, no duplicates, no guessed email addresses.
That is the moat. Not the model, the rules. Anyone can call an API. The value lives in the judgment about what to throw away.
A sample from the rulebook:
- ✅ Keep, a live role on the company's own job board that names the tool in use
- ✅ Keep, every claim carries a verbatim quote plus a specific, working source URL
- ❌ Reject, consulting partners, agencies, and vendors who build on the platform
- ❌ Reject, "Salesforce" appears only in their product or an integration, not their own internal use
- ❌ Reject, competitors, look-alikes, and anything already on our master list or a prior drop log
- ❌ Reject, stale evidence with no current signal to corroborate it
The engine, in four stages
Once the rules existed, the rest became an assembly line the AI runs on a schedule. Four stages, in order:
- Discover. Rather than scrape blindly, the system searches the open web and reads companies' own job boards and tech stacks for live proof the tool is in use. Every candidate arrives with a citation attached.
- Qualify. Each candidate is run against the full rulebook. Anything that fails is dropped, and the reason is written down. A company with only a stale case study and no current signal doesn't make it through.
- Enrich. For the survivors, we find the operators and decision-makers with verified contact details, capped to a set budget each day so it stays deliberate instead of becoming a firehose.
- Personalize. The step most automation skips. Every message is written for one person, grounded in a real signal about their company, not a mail-merge template with a first name dropped in.
The moat was never the model. It was the rules.
Why it doesn't invent a pipeline
The failure mode of AI research is confident fabrication: a tidy list that looks authoritative and is quietly wrong. We built the guardrails to make that hard.
Every company we keep carries a verbatim quote and a working link. A controlled vocabulary keeps the data clean instead of drifting into forty spellings of the same label. There's a hard daily cap, a running audit of anything the system dropped, and one rule we repeat constantly: ten verified companies beat fifteen shaky ones. When a check is uncertain, the answer is to drop, not to guess.
A human still reviews the output. But the human is reviewing evidence, not starting from a blank page.
What it feels like to run
In practice it's a small team of research agents, each working a different angle, every morning, all of them reporting back with sources. What lands is not a mystery list, it's a repository where you can click any row and see exactly why it's there.
The real lesson has almost nothing to do with sales. AI didn't replace our judgment. It let us write our judgment down once, as rules, and then run it at a scale a person never could, without lowering the bar. Encode the expertise. Automate the throughput. Keep the standard. That's the pattern we come back to across everything we build.
This is how we find the people we build for. It's also, more or less, how DealPulse thinks about deal risk inside Salesforce: watch the real signals, trust the evidence, and flag what actually matters before it costs you the quarter.