Our own inbound leads had no automated follow-up — each inquiry was replied to manually or missed.

selfx's own contact and booking forms generated real inquiries with no automated follow-up. Each one was either answered manually — a 30-to-60-minute lag while someone happened to check the inbox — or it slipped through entirely on a busy day. The lead was already paid for by the time it arrived; losing it to a slow reply is the most expensive way to lose a customer.

A complete lead-response engine, wired into the live site. A new inquiry hits a Next.js API route, is validated and written to Postgres with a unique reference, and immediately pings the owner on Telegram (this part is live now). On top of that runs a productized engine (pure decision core, dependency-injected, unit-tested), triggered asynchronously once the lead is saved and switched on when the Claude and email keys are connected: an instant AI reply bounded to a per-vertical whitelist of approved claims, a multi-turn qualification loop that scores lead fit from 0 to 100 and tags the lead, a deterministic opt-out check that runs before the model (genuinely fail-closed) for compliance, and a follow-up cadence drained by an authenticated cron job that stops the moment a lead replies or books. Four vertical presets ship out of the box — consulting, beauty, dental/repair, real estate — the exact speed-sensitive ICP. First-response time is instrumented per conversation (stored as a timestamp; latency is computed from it).

Running live in production. Every inquiry is written to Postgres and pushed to the owner's Telegram within seconds — there is no longer an unwatched inbox where a lead can sit. The full AI reply, qualify, and follow-up engine is built, wired to live routes, and unit-tested across every vertical preset and the inbound path; it activates the moment the Claude and email keys are connected. Because first-response time is already instrumented per conversation, the real before/after number gets measured from day one rather than guessed — so we are not printing a metric we have not measured yet.