The problem with traditional lead scoring
Most sales teams rank their leads by "interest" or "fit": fields filled in by gut feel, during a 30-second call or by reading the prospect’s email. The result is predictable: half the good leads go cold waiting for a call that never comes, and half the team’s time is spent chasing leads that were never going to close.
AI lead scoring is not magic. It is applied probability. Every lead that enters your CRM receives a score from 0 to 100, calculated from what the model learned from your historical deals: which ones closed, which ones fell through, which patterns repeated in each case. The system stops guessing and starts calculating.
How we build it
We don’t arrive with an off-the-shelf model. Every project starts with your history, not a generic template.
1. We audit your history
Before training anything, we need at least 200 historical deals in your CRM. If you have fewer, we’ll tell you: the model won’t learn anything you can’t already see yourself. If you have them, we analyse them: which signals correlate with closing, which signals seem important but aren’t, which leads were discarded prematurely.
2. We train the model
We use gradient boosting (XGBoost or LightGBM depending on data volume). We don’t use LLMs in this layer: these models are faster, more explainable, cheaper to serve, and for tabular classification they are state of the art. Each prediction tells you why — which signals influenced the score and in which direction.
3. We integrate it into your CRM
The score appears as one more field on each lead. HubSpot, Salesforce, Pipedrive, Dolibarr — wherever your CRM is, that’s where the score goes. No third-party plugins, direct integration via API. Your team doesn’t learn a new tool: they see one more number on the record they already know.
4. We set the cut-offs
Lead with score > 70 → immediate contact from the sales rep. Score 40-70 → automatic nurturing (predictive email, retargeting). Score < 40 → outside the active pipeline or reclassified for review. We fine-tune the exact figures with you during the first 4 weeks by observing how the model behaves in production.
What changes when this works
You notice it in three places fairly quickly:
- Your sales team stops chasing leads that were never going to close — reclaiming hours and energy lost to false positives.
- Good leads don’t go cold waiting their turn. The system tags them instantly and moves them to the top of the queue.
- Pipeline review meetings shift from discussing gut feelings ("I’ve got a good feeling about this one") to discussing numbers ("score 82, high fit, two visits to the pricing page").
Companies we have done this with report +25-40% in close rate on qualified leads and a 60% reduction in response time to top leads. We do not promise those numbers — we have seen them, but they depend on your CRM and your team. We measure it honestly.
When we do NOT recommend lead scoring
Before the first quote, we’ll tell you:
- If you have fewer than 200 historical deals in your CRM. The model doesn’t have enough signal.
- If your sales cycle is ultra-short (<48h). Scoring adds no value at that speed.
- If you sell to a closed list of key accounts. What you need here is ABM, not scoring.
- If your sales team already closes 50%+ of the leads they touch. Their instinct works — no additional layer needed.
In any of these cases there are better levers. We recommend which one and where to focus the effort.
Retraining is not optional
A model trained in January loses accuracy by July. The market changes, your product changes, your competitors change, your audience changes. Every quarter we re-train the model with new data, compare metrics (AUC, precision/recall, calibration) and adjust the thresholds if needed. It’s in the contract. This is not a set-and-forget — it’s an ongoing service.
Stack we use
- XGBoost / LightGBM as the base model, with stratified cross-validation.
- Python (scikit-learn, pandas, MLflow) for the features pipeline and model versioning.
- Direct integration with your CRM via REST API. No third-party plugins, no Zapier in between.
- Custom dashboard in Metabase or Looker Studio so you can see the scores, the distribution, and model drift.
We start with a diagnostic session
Before quoting the full setup, we run a 90-minute session. We look at your CRM, your historical deals, your pipeline together. We leave with an honest recommendation: whether lead scoring is right for you today, or whether it makes sense to start with something else (commercial audit, AI CRM, predictive email).
We don’t charge for that session. If you’re interested, get in touch.