How to Build an AI GTM Brain on GLM-5.2: A Signal-Based Outbound System You Run Yourself

Open original

This article walks through a DIY outbound sales system using GLM-5.2 to spot fresh company signals, score and rank accounts, draft personalized openers, and self-evaluate before sending. It covers setup, dry-runs, free public signal sources, scoring logic, trigger-grounded messaging, an eval gate, and a learn loop on a schedule.

  • 142 signal-qualified leads produced 19 meetings in two weeks (12% reply rate), driven mainly by funding news and site visits, while social likes were skipped as low-signal noise.
  • The system gates itself before going live: it must prove hot leads outperform cold against a golden set, and drafts without a fresh trigger get killed rather than sent.
  • Runs on GLM-5.2 via API for a few dollars a month, pulling free public signals (job boards, funding news, competitor wins) plus first-party site visits, fired daily by a single cron job.
  • Output is a ranked daily list with a one-line "why-now" per account, letting a human send/edit/skip each in under two minutes.

Sortlist’s AI GTM Brain flips the usual outbound playbook by starting with real-time signals, not a bulk list. You point GLM-5.2 at Anthropic’s API (or swap in a free model), and the system costs you only a few dollars a month. It watches public job boards, funding announcements, competitor wins, first-party site visits and, if you choose, your own Sortlist demand feed. When a “why-now” trigger pops up—a fresh round or a reposted role—it scores that account against its full history and your ideal customer profile, then spits back a “send” or “skip,” complete with a quoted one-line opener. Every step runs itself, but you stay in control. First you test on sample data until the gate passes. Then before any live send, the brain grades its own accuracy against a golden set and halts if it can’t prove hot leads outperform cold. Drafts that don’t tie to a fresh trigger die. Sequences stay short—no padding nine follow-ups to force relevance. You log outcomes, the model retrains its bucket weights, and a single cron job fires it off each morning. On Sortlist’s pipeline, this method booked 19 meetings in two weeks from 142 signal-qualified leads—a 12% reply rate. Funding news and site visits drove most calls; social likes were skipped, which kept noise down. Every morning you get a ranked list with a precise why-now under each name, then choose send, edit or skip in under two minutes.

Ask about this article