B2B lead scoring with AI: how to build a model that actually predicts revenue
Most lead scores predict nothing. The five-input model that actually correlates with closed-won revenue — buildable in a weekend.
Most B2B lead scoring models are astrology in a spreadsheet. Points for downloading an e-book, points for visiting the pricing page, threshold-based MQL. None of it correlates with revenue in any statistically meaningful way.
The reason is that lead scoring is usually built by marketing, in isolation, without joining to closed-won data. You end up scoring engagement, not intent, and rewarding tire-kickers.
The rebuild is straightforward. Take the last 12 months of closed-won opportunities. Take the last 12 months of closed-lost. Find the five inputs that separate them best. That's your model.
In our practice, the five inputs almost always include: company size (proxy for ACV), title of first touchpoint, source of first touchpoint, signal presence (funding, hiring, tech installs), and time from first touch to a qualification question.
AI makes the scoring dynamic instead of static. Instead of 'this lead has 87 points,' it says 'this lead resembles won deal patterns in 72% of features.' The confidence interval is the actionable output, not the number.
The routing changes too. Top 10% of scored leads go to a senior AE within 5 minutes. Middle 60% get an AI-agent sequence. Bottom 30% get nurture. Most teams give everyone the same treatment and lose the top 10% to slow follow-up.
Recalibrate the model quarterly. Buyer behavior changes. Your product changes. A model that's more than 90 days stale is drifting into astrology again.
The measurement is model precision at the top decile: what percentage of your top 10% scored leads become opportunities? If it's below 25%, the model is worse than a fit-only filter. Rebuild it.
Frequently asked questions
RevOps — answered
- Do I need machine learning or is rules-based fine?
- Rules-based works up to a few hundred leads a week. Above that, ML meaningfully improves precision.
- What tools should I use for AI lead scoring?
- Most modern CRMs and MAP platforms have native predictive scoring now. Roll your own only if you have >5,000 leads a month.
- Should I score cold outbound the same as inbound?
- No — the input weights differ. Cold outbound scoring emphasizes fit and signals; inbound emphasizes behavior and source.
- How often should I retrain the model?
- Quarterly, with a full rebuild every 12 months. Buyer behavior shifts faster than most teams update models.
Growth Broker editorial
Filed under revops