The ICP filter is the deliverable. The list is the byproduct.

Drop Water builds a self-serve beverage station — still and sparkling water, sixteen flavors, four functional boosts, ten bag-in-box slots, 4,720 servings a machine, restocked in ninety seconds by whoever already works there.

The product was not the problem. Roughly two-thirds of pipeline arrived inbound or by referral, which is a compliment and a ceiling at the same time. There was no written ideal-customer profile, no way to tell a qualified prospect from an unqualified one before paying for the contact, sequences written against assumptions nobody had checked, and outbound running from the company's primary domain. An earlier campaign had already gone to market: 164 sends, one meaningful reply, and nothing built to catch it.

That last line is the origin of half of what follows. A reply rate is a copy problem you can fix in an afternoon; having nowhere for a reply to land is a systems problem, and it wastes every send that came before it.

So the work started at the filter rather than the list. Contact data decays in weeks; the filter that produced it does not. And most of the criteria that actually qualify a prospect — owner on the floor, pricing autonomy, franchise lock — are not visible in any contact database, which means the filter has to be built and measured rather than bought.

An ICP filter with its coverage measured, not assumed. 2,855 company records probed across two personas, which established that roughly 80% of correctly-identified targets have no reachable buyer in the contact database the stack was built on. The hypothesis that larger operators would have better coverage was tested and disproved. That single measurement redirected an entire vertical's sourcing strategy before the budget went into it.

A prospect list built from first-party sources. 546 domains in, 411 sendable contacts out — of which 196 carried a real named human, 47.7% against the 10% the handoff projected. The difference came from a manual pass over each business's own website rather than another database seat. Every record ships with an audit file carrying provenance per field and a rejects file showing what did not make it.

An account layer keyed to survive the CRM. 411 accounts, 411 distinct domains, clean one-to-one, zero collisions. This exists because the CRM associates contact to company by email domain — which would have permanently orphaned the 107 contacts, 26% of the list, using a personal email address. In small-business categories, a quarter of real owners use Gmail.

Two production sequences and a twelve-field personalization layer, with an A/B/C test designed in, and a reply-handling model built from the post-mortem of the campaign that failed — categorisation and notification, with a human writing the response, because an objection is a conversation and a templated answer wastes a reply that took nine touches to earn.

This engagement is infrastructure-complete and outcome-early. As of the last logged entry, both sequences were built and sends had not started, so there is no reply rate to publish and we are not going to invent one.

What is provable today is the discipline, and it is the more useful story anyway:

An independent quality gate scored a finished sequence at 73.6 against the drafter's self-scored 84.2 — roughly ten points of inflation per touch — and found the required opt-out missing from 8 of 10 emails in a set that had self-certified it present. Nothing has shipped since without an outside pass.

An import returned 145 successes and 169 failures. Reading the failure file rather than the success count turned up something worse than the error: nine company names that would have shipped inside a live subject line — “A note for whoever runs coming soon” among them. All nine corrected, and a standing check written: print the assembled sentence and read it, do not just check the field for non-emptiness.

Two enrichment shortcuts were tested and rejected on the numbers rather than quietly shipped — deducing first names from email local parts measured 37% precision against 115 independently-verified names.