A build week. Most of it converts to demand next week: the lists go to outreach, the SMS system flips live, and the Learning Centre work goes to a ship decision.
We ran the enrichment pipeline over three lead lists this week. The headline file is the SimplePractice provider directory, cross-checked against the federal NPI register.
| List | Contacts | Phone | Own website | LLM cost | |
|---|---|---|---|---|---|
| SimplePractice directory86% confirmed against the gov NPI register | 39,010 | 99%64% net-new (24,985 we added) | 60%5,327 personal · 12,305 practice | 45% | $2.42 |
| Psychology Today · Canada | 3,781 | in source | 48% → 54% | 58% → 65% | $0.08 |
| US health multi-directory | 5,840 | 41% → 64% | 0% → 26% | 0% → 47% | $0.45 |
An interim pass looked even better, until parallel reviewer agents found a parsing bug counting tracking IDs and timestamps as phone numbers. The figures above are the corrected, re-verified set.
$2.95 total LLM spend; all-in with prepaid search credits is roughly $50 to $80 (estimate). Not in outreach yet: emails get a verification pass first, then the SDR-vs-cold-email channel call. Cold email would run on a separate domain, never production Customer.io.
About 515 US signups a month arrive with no phone number. The same techniques that enriched the directories now run against them, so sales can actually reach the people who just signed up.
A 21-node n8n workflow on our n8n cloud, reading real signups and writing nowhere. Verified against the local prototype: 32 of 32 leads scored identically. Hardened against server-side request attacks before it ever touches production data.
Directory rows come anchored to a real person; a signup is just a name and an email. New approach: a free, deterministic first pass against the federal NPI register (reverse phone-to-identity), with the LLM now allowed into identity matching under strict rules.
Current strike-rate estimates: roughly 45% on completed signups, 25 to 30% on abandoned. Auto-accept yield is still near zero until the matching weights are re-tuned on a golden set. That is next; nothing goes live before it.
E-learning engagement keeps showing up as a standout signal, but only 7 to 8% of dashboard clickers ever open Learn. This week's pass makes the centre worth opening.
Role and goal entry paths on the home page (3 roles, 5 goals), a 75-question knowledge-check bank across every lesson, per-course wrap-up quizzes, printable course checklists, and a video layer with the first five videos slated.
Analytics events wired for quizzes, videos and checklists, so we will know what people actually use. All 47 automated checks and the brand gate pass; the 17 pricing claims in course content re-verified, one stale FAQ fixed.
A working prototype, not yet live. Ship decision comes after review and video recording.
Every reply to our SMS campaigns has been landing in a demo bot instead of a human. 959 replies and counting. This week we built the system that fixes it.
An LLM classifier inside Customer.io sorts every reply: opt-out, support issue, needs a human, or no action. It only classifies. Every customer-facing SMS is pre-written copy, so it can never improvise an answer. Support issues route to CS, the rest get a booking-link acknowledgement, and every real reply creates a due-today activity for sales.
A catch-up broadcast to the resolvable backlog: your text never reached a human, that was our fault, it's fixed, here's a direct line. Opt-outs and dead numbers excluded.
The system is running in dress rehearsal right now: real replies flow through the routing, but everything outbound queues as drafts. After a clean 4-hour read, it flips to real sends.
Three tiles became seven, an anchor-lag bug fixed. The clean finding: conversion climbs monotonically with activity depth, 2.4% for one-active-week workspaces up to 12.1% for four of four. Follow-through is wobbling; the verdict on that lands over the next month.
The contract and intake reminder ladders (live, first ~40 sends out) were sending three byte-identical nudges. Replaced with 6 escalating templates, QA'd clean, written straight to production.
Re-ran the cohort analysis: Feb to Apr $1 redeemers drop from 88% to 71% survival between days 100 and 130, and ~181 accounts roll off in the next 10 weeks. A retention play for that window is the obvious next move.
Day-7 retention re-read: +2.1pts (12.1% to 14.2%) for the new dashboard. Not yet statistically significant, roughly 93% probability it's real. Still gathering power before we call it.
48,631 enriched contacts in hand, two live channels tightened, and three systems one review away from switching on.