AI for marketing
Predictive lead scoring for B2B teams
What a predictive score means, the data it learns from, how to judge its accuracy, and how to get sales to use it.
Key takeaways
- A predictive score is a probability learned from your closed leads.
- Clean won and lost outcomes in the CRM matter more than extra signals.
- Test the model on recent deals it did not train on before launch.
- Show sellers the reasons behind a score, not only the number.
- Retrain on a schedule and watch for segments the model undervalues.
Predictive lead scoring uses a machine learning model trained on your past won and lost leads to estimate how likely each open lead is to convert. Unlike points-based scoring, nobody sets the weights by hand. It works when your CRM records outcomes cleanly, and it needs testing against recent deals before sales relies on it.
What is predictive lead scoring?
It is a way to rank open leads by how likely they are to convert, using a model trained on the leads you have already won and lost. The model finds the patterns in your history and applies them to new leads.
The output is a probability, not a verdict. HubSpot's likelihood to close property, for example, estimates "the probability that your open contacts will close as customers within 90 days."[1] Microsoft's Dynamics 365 Sales scores each lead on how likely it is to convert to an opportunity and groups leads into grades A to D.[4]
Definition
A predictive lead score is a model's estimate of the chance that a lead reaches a defined outcome, such as becoming an opportunity or a customer, within a defined period. The outcome and the period are what make the number meaningful, so write them down before anyone acts on a score.
How does it differ from points scoring?
In points-based scoring, people pick the signals and set the weights: a meeting request adds points, a student email address removes them. The rules are easy to read but drift away from reality as your market and channels change.
A predictive model learns the weights from outcomes instead. It can pick up combinations that people would not think to write as rules. The cost is less control and, with some tools, less visibility. Many teams keep simple fit rules, such as country and company size, as a filter and let the model rank leads inside them.
What data does the model need?
Outcomes first. A model can only learn from leads that were clearly marked as won or lost, qualified or disqualified. Microsoft's Dynamics 365 Sales will not build a lead scoring model until the organization has "created and closed at least 40 qualified and 40 disqualified leads" in the chosen training period, which ranges from three months to two years.[2] That is a floor for one product, not a target for a custom model.
Then signals. HubSpot lists the inputs its model uses: website and email engagement, CRM activity and contact fields, firmographics for the contact's company, and details about your own business and account.[1] In practice the useful signals for B2B are usually:
- Fit: industry, company size, region and role, matched to your ideal customer profile.
- Engagement: pages viewed, forms submitted, emails replied to, meetings booked.
- Sales activity: whether and how fast someone followed up.
If your CRM stages are inconsistent or leads are closed without a reason, fix that before you buy or build anything. Our guide to closed-loop attribution from CRM to ad spend covers the same data hygiene from the marketing side.
Is your CRM ready for predictive scoring?
A free growth diagnostic reviews two or three time-heavy workflows, your data readiness, quick wins with current tools, and data and policy risks, and ranks your top 3 use cases by value and effort. Your top 3 findings and a 30-day fix plan in writing within 2 business days.
How do you check a model before launch?
Test it on recent leads it did not see in training. Microsoft trains on 80% of closed records and validates on the remaining 20%, "which consists of the most recent records."[3] That is a sensible pattern for any model, because it tests the model on the period closest to the leads it will score.
Look at more than one number. Microsoft reports accuracy, recall, F1 score and the area under the curve (AUC), which is the probability that the model ranks a randomly chosen qualified lead above a disqualified one.[3] Accuracy alone can mislead when most leads never convert, so compare how the top group of scored leads converted against the rest.
Then run a short shadow period: score leads without changing routing, and check whether the high scores really converted more often.
Illustrative sample
A hypothetical software company scores inbound leads for a month without changing who gets called. At the end, the sales lead compares meetings booked from the top fifth of scores with the rest, reads ten high-scoring leads that went nowhere, and only then routes the top grade to a fast follow-up queue.
How should sales use the score?
As a way to order the queue, not as a gate that hides leads. Grades or tiers are easier to act on than raw numbers, and both HubSpot and Microsoft group scores into tiers or grades for that reason.[1][4]
Reasons matter for trust. Microsoft's lead score widget shows the top five positive and negative reasons behind each score.[4] HubSpot, by contrast, says "it's not possible to know exactly how each input contributes to a contact's score."[1] Sellers are more likely to use a score they can explain to themselves. Pair the score with clear routing and follow-up rules; our article on LinkedIn lead gen form routing shows one way to get leads to the right person fast.
How often should the model be retrained?
Often enough to follow your market. Buyers, channels and products change, and a model trained on old outcomes keeps ranking leads the old way. Microsoft offers automatic retraining every 15 days[2] and refreshes scores for updated leads every 24 hours.[4]
Beyond the schedule, review:
- Missed leads: low-scoring leads that sales picked up and won anyway.
- False alarms: high-scoring leads that were disqualified, and why.
- Inputs: fields that are often empty or changed meaning, which can distort the model.
What are the risks?
The biggest one is that the model copies your past. If most past wins came from one segment, it will score that segment highly and may undervalue a new market you are entering. Keep a human path for leads that matter strategically but score low.
Second, data protection. Scoring uses personal data, so your privacy notice and lawful basis should cover it, and the inputs should be limited to what you need. Our article on AI governance for LLM training covers data controls that apply to scoring too, and normalizing unstructured data for LLMs covers cleaning the records models learn from.
Where should a B2B team start?
-
Define the outcome and the time window the score predicts.
-
Clean the won, lost and disqualified records for the last one to two years.
-
Try the scoring built into your CRM before building a custom model.
-
Test on recent records, then run a shadow period before routing changes.
-
Agree how sales will use each grade, and review results on a set schedule.
Our AI marketing service covers these steps as part of our AI consulting work.
Sources
Written under our Editorial and AI Policy. Sources checked 19 Oct 2026. Reviewed by DGo Digital on 19 Oct 2026.
FAQ
Questions about predictive scoring
In rules-based scoring, people choose the signals and assign points to each one. In predictive scoring, a model learns the weights from leads that were won or lost in the past and outputs a likelihood of conversion. Both need clean CRM data, and many teams run them side by side while they compare results.