My best-fit customers aren't in any database you can buy. An exercise in thinking outside of the box.

My ideal customer isn't in any database you can buy. Sit with that for a second — it breaks the whole GTM playbook. Small and mid law firms in Brazil: local, WhatsApp-run, invisible to the firmographic backbone every enrichment tool resells.

I could scrape a list, of course — but a list of names is worth nothing if I can't trust a single field on it. My idea was simple: if waterfall enrichment is basically putting several data sources against each other to see what's actually true, it could be built from the ground up.

Here's the honest limit most enrichment decks skip: a list can tell you who exists. It cannot tell you who's a fit. In a market with signal — funding, hiring, tech stack — you paper over that gap with intent data. My market has no such gift: small Brazilian law firms are nearly invisible to the databases, and the ones with a website are the exception.

So I drew a hard line about what enrichment is allowed to conclude. It builds a picture — identity, rough revenue, practice area, real name, live channel — and writes a briefing good enough to open with a real question. It does not rule anyone in or out on fit, because it can't: a solo lawyer billing R$2k who wants to close more contracts might be perfect; the same R$2k split across partners isn't mine; R$40k is a coin toss until they talk. The conversation qualifies. Enrichment just earns the right to start it.

(There is stronger signal — actual intent and pain — but it lives where the ICP already communicates: strong, and far scarcer. I'm building for that now. Another story, another time.)

The turn was realizing I already had the data — I just hadn't been reading the right records. The raw material is public, if you know the law. In Brazil, whois for a .brdomain is open by statute, and the federal company registry — CNPJ, partners, share capital — is a matter of public record. So the pipeline reads what the vendors resell and the operators ignore: Jina for the site, registro.br for the domain's owner, Receita for partners and capital, Google for reviews, the page itself for a Meta pixel or GTM tag, Instagram for bio and reach. None of it fires blindly — no site means no pixel, no CNPJ means no Receita — and each source is there to fill or correct what the one before it missed. That's waterfall enrichment, except I own every stage.

And the sources overlap on purpose. Jina pulls an email, a phone, a name off the site; the whois pulls an email, a name, and the CNPJ; Receita takes that CNPJ and returns the partners, the capital, the practice area. The same fields surface two or three times from different places — and that overlap is the whole game: a CNPJ from the whois unlocks the record that confirms the rest, and a value that shows up in more than one source is a value I can trust. Then an LLM reads every source at once and reconciles them into one record — classifying the firm and resolving each field, preferring what more than one source agrees on.

Here's the smallest decision on the page, and one of the ones I'm proudest of. The name I'm after isn't just any name — it's the decision-maker's. A cold WhatsApp addressed to the partner who actually runs the firm, by name, lands nothing like a generic intro to the institution: one reads as someone who already knows who's in charge; the other as a stranger who found a number and is hoping to be passed along. Get the name wrong, or leave it blank, and the lead knows in one second they're talking to software. So the name is the one field the system is never allowed to invent.

With intent off the table, the next best question is cheaper: which of these firms even looks like it can afford to fix the problem? The score answers that, and nothing more. It's a plain additive model — every signal the enrichment found adds points. A professional site, not a drag-and-drop one. Ad spend already running. A Meta pixel on the page. Reviews in the range a mid-sized local practice tends to sit in. Real share capital and a sane number of partners at Receita. A high-ticket practice area — corporate, tax, real estate. Each is a proxy, and I treat it like one: no single signal means much, but stacked, they sort a firm into three tiers on two fixed thresholds. It's a read on capacity, not fit.

Here's where it would have been easy to be clever, and wrong. The obvious move is to let the score gate: qualified firms get worked, the rest get dropped. I built the score and then refused to let it do that. It writes to a note on the CRM record — it colors the lead, it tells the human where to look first — and it decides nothing about who gets contacted. The reason is the discipline the whole system runs on: I hadn't validated this market, so a threshold I invented is just a guess wearing a number, and gating on it is the operator's mistake at smaller scale. Worse, the score reads capacity, and fit lives in the conversation — the firm that looks marginal on paper is exactly the solo lawyer who turns out to be perfect once he talks. So only facts gate, never the score. No WhatsApp, no contact — you can't work a channel that isn't there. No decision-maker I can name — I won't open blind. A firm too big to be my market — a different sale than the one I built this for. Everything else earns a conversation, and the conversation decides.

I built this as a cold-call briefing: a profile by digital maturity — no site, site but no ads, already advertising — the pain most likely to be true, a hook, and three NEPQ questions pointed where it counts (cash tied up, time bled, the cost of standing still). Then I never made the calls — solo, there was no time. But it didn't go to waste; it became the seed of something better. It's what whoever takes the booked meeting walks in with, and it doesn't stay static: the WhatsApp BDR enriches it live, folding in what the lead actually says on the thread — the real pain, in their words. So by the time there's a demo, the briefing is two layers deep: the hypothesis enrichment guessed, and the signal the conversation confirmed. Enrichment writes the first draft of the meeting; the conversation finishes it.

The entry gates decide who gets in; a second layer verifies the channels before anything ships. Every email is checked at the SMTP level by Reacher — if it can't be confirmed deliverable, it's wiped, not saved, because an unverified address isn't a lead detail, it's a bounce waiting to poison your own sending domain. Every phone is checked against WhatsApp itself — a number without it is a dead end on the one channel that matters here. A record ships with a real, reachable channel or not at all — because everything downstream is only as healthy as the data it runs on.

The honest failure mode isn't in the code — it's in the data. Public record is only as current as the last person who updated it. A domain's registrant can be the accountant who set it up, not the lawyer who runs the firm. A partner listed at Receita can have left years ago. The cross-check catches a lot of this — a name that shows up in one source and nowhere else gets treated with suspicion — but it's mitigation, not a cure. Some records still go out thin, or subtly wrong.

And I chose to let them. In a market this quiet, discarding every imperfect record means discarding the market. So the pipeline tolerates a noisier top of funnel than a bought list ever would — an incomplete picture, an occasional wrong field — because the conversation is the real validator, and a lead I can reach is worth more than a database entry I can't. The trade-off, stated plainly: I traded clean data for reachable data, and made the conversation clean up the difference.

This isn't the page with conversion metrics — those live in the conversation. What this build is measured by is cheaper, and for a GTM hire more telling: