Signal-based outbound rebuild
Outbound at a B2B SaaS revenue org. Reply rate sat at 1.2% and meeting-booked at 0.4% after two outsourced vendor cycles. The team thought they needed more headcount. The real constraint was upstream.
The vendor produced noise that looked like activity but never tied to pipeline.
Two-phase rebuild.
Phase 1 was a qualified reply rate classifier. An 8-category LLM separated qualified replies from routing, auto-responders, and noise. The optimization target shifted from reply rate to qualified reply rate.
Phase 2 was a signal-based front end. Two tracks. Personal-level identified visitors went straight to scoring. Anonymous company-level visitors triggered a 7-stage Clay enrichment waterfall that surfaced up to 3 ICP-fit people per company. A 100-point composite scored fit and intent. Above 75 entered the qualified-reply-rate-instrumented outbound. SDRs got Slack pings when high-fit visitors landed on site, with agent-generated prep files surfacing in seconds.
Stack ran on Salesforce, HubSpot, Clay, RB2B, custom MCP servers, and an eval framework that gated every agent output before ship.
- ✓Reply rate moved 1.2% → 3.8%
- ✓Meeting-booked moved 0.4% → 1.6%
- ✓Monthly funnel at steady state. 30K visitors, 9K identified, 1K-1.5K ICP fit, 150-300 above 75, 30-60 meetings booked
- ✓Counterfactual headcount, 3-4 FTEs (researcher, scoring lead, 2 SDRs, sales prep)
- ✓Timeline from proposal to production, 1-2 months
- ✓End-to-end from webhook to scored and routed opportunity, under 3 minutes
Instrumentation before optimization. The leverage lives upstream in the signal capture and downstream in the cost-curve change. The personalization layer is the middle of the system.
Without the front end, you optimize what the vendor was. Which is funding noise.