REVOPSlead-scoring · advanced
AI lead scoring that beats the rules-based one
Last reviewed: 2026-05-23 · saves ~saves SDRs 10 hr/week on dead leads/run
Our take
Most lead scoring is points-based and wrong: '+10 for opened email, +20 for visited pricing page' — calibrated by guess. AI scoring uses actual conversion data + behavioral signals + firmographic match in a regression-validated model. Result: SDRs work the top 20% of leads, not the most-recent 100.
Tool stack
Steps
- Export 12+ months of leads with outcomes: which converted to opportunity, which to closed-won, which never engaged.
- Use Claude to identify predictive features beyond the obvious (page visits, email opens).
- Train a logistic regression or simpler scoring model in a spreadsheet — don't overengineer.
- Deploy as a Salesforce/HubSpot formula field. Update weights quarterly.
- Compare AI-driven score routing vs old rules-based for 60 days.
Prompts
Identify predictive lead features from historical conversion data · Claude Sonnet 4.6
You are a marketing operations analyst. I'm sharing 12 months of lead-level data with conversion outcomes (lead → MQL, lead → opp, lead → CW). For each feature column (e.g., title contains 'VP', source = LinkedIn ads, last activity in 7 days, company size 50-200, intent topic match, etc.), tell me: 1. Conversion rate when feature = true vs false 2. Lift (relative to baseline) 3. Statistical significance (back-of-envelope; flag if < 30 conversions) 4. Recommended weight in scoring model (0-30 points) Output a markdown table, sorted by lift descending. At the end, recommend a final scoring formula using the top 7 features, with explanation of why each made the cut. Constraints: - Features with < 30 conversion events flagged as noisy. - Avoid features that are just proxies for 'this lead converted' (causation-direction). - Be skeptical: if data shows nothing useful, say so.
Pitfalls
- Don't let the model 'discover' that big-company leads convert better — that's a known bias, not insight.
- Watch for survivorship bias: leads that didn't engage may have been routed away by the old model.
- If your historical data is < 500 leads, AI scoring is noise. Use heuristics.