AI Agent · B2B Sales
Outbound Engine
A self-improving AI agent that sources, qualifies, and writes outreach for high-intent B2B leads.
Overview
Outbound Engine is an AI-powered outbound prospecting system that autonomously finds and qualifies high-intent B2B leads from public web signals, then drafts personalized outreach openings for each one. It runs as a self-improving agent: each run sources fresh prospects, scores their buying intent, filters out noise, and writes the results to structured pipelines ready for outreach. Over time it learns which search strategies and message angles work and retires the ones that don't.
The challenge
Founder-led and lean teams waste hours manually trawling LinkedIn for prospects, and generic lead lists are full of agencies, recruiters, stale posts, and ICP mismatches. The goal was to replace that manual grind with a disciplined agent that only surfaces genuinely high-signal buyers, for example a founder publicly hiring their first engineer with traction metrics, while enforcing strict deduplication and recency rules and explaining its reasoning per lead so a human can trust and act on each one.
What we built
- A signal-driven lead-sourcing engine that queries public, indexed surfaces with a library of tuned search queries, prioritizing company-homepage and founder-activity discovery chains over generic search results.
- Multi-factor qualification and intent scoring, giving each lead a persona tier, an intent score backed by explicit signals, and a confidence score, with hard filters for competitors, geography, recency, and duplicates.
- A self-improving learning loop with a persistent log that tracks cumulative stats, per-query performance (keep or retire), persona performance, a mistakes log, and an evolving rule set the agent applies and refines on every run.
- Opening-message generation with A/B tracking that measures message-template performance by response rate to learn the most effective angles.
- A structured, auditable output pipeline emitting dated CSVs with persona, intent, signals, and recommended opening per lead, plus run JSON, a duplicate tracker, query-performance data, and an agent memory file capturing next-run focus.
Results
Across roughly two dozen automated runs the system sourced well over a hundred prospects and qualified more than half against a deliberately high bar, with per-lead reasoning that keeps the pipeline auditable. Lead quality improves run over run as learned rules retire noisy queries and favor stronger signals such as funding-plus-hiring activity. It dramatically reduces manual prospecting time per qualified lead and leaves behind a compounding knowledge asset, each run ends better tuned to the ideal customer profile than the last.
