Predictive lead scoring sells well: a model trained on your won/lost deals tells you who’ll convert. In practice it underperforms a well-tuned rule-based score until you cross 5,000+ closed deals of training data. Here’s the honest review.
Where predictive wins
When you have over 5,000 closed deals across 18 months, varied across industries and deal sizes, and your won/lost ratio sits between 15% and 40%, the model finds non-obvious patterns. It will surface combinations your rules missed (e.g., contacts who visited the integrations page twice but never the pricing page convert at 2x).
Where predictive loses
If your training data is under 1,000 won deals, the model overfits to noise. You’ll see weird ranking (a tiny prospect ranked above an enterprise account) and reps will lose trust in week two. Stick with rules until you have volume.
The data hygiene tax
The model is only as good as your closed_won and closed_lost labels. If reps mark deals as closed_lost for “wrong reason” (timing, no budget, dead lead), the model learns the wrong signal. Audit your closed-lost reasons and force a clean taxonomy first.
Hybrid is the right answer
Use predictive for ranking among the top quintile and rules for everything else. The rules filter out obvious junk; the model decides who the SDR calls first. Configure your routing as:
Step 1: Filter contacts where rule_score >= 50
Step 2: Sort by predictive_likelihood DESC
Step 3: Assign top N to SDR queueRecalibrate quarterly
The predictive model retrains automatically but your business shifts faster. New ICPs, pricing changes, market conditions. Force a manual retrain every 90 days and review the top contributing features. If “lead_source = Old Trade Show” is still in the top 10, your model is stale.
Don’t show the score to the prospect-facing team without explanation
Reps don’t trust black boxes. Always pair the predictive score with the top three contributing features in the contact card. HubSpot exposes these via the model insights panel; surface them.
What to do this week
Pull your closed-won and closed-lost counts for the last 18 months. If you’re under 1,000 won, defer predictive and double down on rule-based scoring. If you’re over, audit your closed-lost reason hygiene before enabling.
