AI Lead Scoring That Actually Predicts Revenue
A B2B SaaS company replaced gut-feeling lead scoring with an AI agent — increasing conversion rates by 34%.
The Problem
A B2B SaaS company with 2,000 monthly leads had a classic sales problem: their CRM lead scores were worthless. Marketing assigned scores based on a static point system — download a whitepaper (+10), visit pricing page (+20), have a .edu email (-50). But the scores didn't correlate with actual revenue.
Sales reps ignored the scores entirely and cherry-picked leads based on gut feeling. Pipeline reviews were debates about which deals to prioritize. Win rates sat at 12%.
The Solution
We built an AI lead scoring agent that:
- Analyzes 40+ signals per lead — not just website activity, but firmographic data, tech stack, hiring patterns, funding rounds, and engagement velocity
- Predicts 90-day revenue probability — not just "hot/warm/cold" but expected deal value and close timeline
- Explains every score — reps see exactly why a lead scored high or low
- Updates in real-time — scores change as new signals come in (email reply, meeting booked, competitor mentioned)
- Learns from outcomes — closed-won and closed-lost deals retrain the model monthly
The Architecture
The scoring pipeline runs on three layers:
Data enrichment layer: CRM data + Clearbit firmographics + website activity + email engagement + LinkedIn data + G2 intent signals.
Feature engineering: Raw signals get transformed into meaningful features. "Visited pricing page 3 times in 2 days" becomes pricing_urgency_score: 0.87. "Company raised Series B last month" becomes budget_expansion_signal: 0.72.
Prediction model: A gradient-boosted model predicts two things: probability of closing within 90 days, and expected contract value. The output is a single composite score from 0-100.
The Results
| Metric | Before | After |
|---|---|---|
| Lead-to-opportunity conversion | 8% | 14% |
| Opportunity win rate | 12% | 18% |
| Average sales cycle | 67 days | 49 days |
| Revenue per rep per quarter | $180K | $242K |
| Rep adoption of scoring | ~10% | 89% |
The key unlock was transparency. When reps can see "This lead scored 82 because they match 4 of your top 10 closed-won firmographic patterns and visited pricing 6 times this week," they trust and use the scores.
Key Takeaways
- Static point systems don't work. The relationship between signals and revenue is non-linear and changes over time. You need a learning system.
- Explain every score. Black-box scores get ignored. Transparent scores get used. Every prediction should come with a human-readable rationale.
- Firmographic signals matter more than behavioral ones. Company size, industry, tech stack, and funding stage predict revenue better than page visits.
- Retrain monthly, not quarterly. Market conditions change fast. A model trained on Q3 data makes bad predictions in Q1. Monthly retraining keeps accuracy high.