Predictive lead scoring uses a machine learning model trained on your past leads, both the ones that became customers and the ones that didn’t, to estimate how likely each open lead is to convert. Instead of points your team assigns by hand, the model finds which traits and behaviors separated your wins from your losses, then scores new leads against that pattern. Sales works the highest scores first.
In HubSpot, predictive scoring appears as two contact properties, Likelihood to close and Contact priority, on Marketing Hub Enterprise and Sales Hub Enterprise. This guide covers how predictive scoring works, how it differs from rule-based scoring, how much data it needs, how to use it in HubSpot, and its limits.
How predictive lead scoring works
Every predictive scoring tool follows roughly the same five steps:
- Collect outcomes. The model needs leads with a known result: converted or not, closed-won or closed-lost.
- Gather the signals. Fit data (industry, company size, job title, location), behavior (page views, email engagement, form submissions, meetings), and CRM activity attached to those leads.
- Train the model. It compares converted leads with the rest and learns which combinations of signals predict a win.
- Score open leads. Each open lead gets a probability or a score, and many tools also group leads into tiers.
- Retrain. As new deals close, the model updates. Microsoft’s Dynamics 365, for example, can retrain automatically every 15 days (Microsoft Learn).
The score is only as good as the outcomes behind it. If junk leads were marked closed-lost alongside real lost deals, the model learns to tell junk from customers and little else.
Predictive vs traditional lead scoring
Traditional (rule-based) lead scoring adds and subtracts points for rules your team writes. Predictive scoring learns the rules from your data.
| Traditional (rule-based) scoring | Predictive scoring | |
|---|---|---|
| Logic | Rules your team writes, such as +10 for a director title | A model trained on your past conversions |
| What it catches | Only what you encoded | Patterns in your data, including ones nobody wrote down |
| Data needed | Works on day one | Needs enough won and lost leads to train |
| Explaining a score | Every point traces back to a rule | Depends on the tool, from top factors to nothing |
| Upkeep | You prune and re-weight rules by hand | Retrains as new outcomes arrive |
| Bias risk | Your assumptions about the ICP | Your history, including segments you no longer target |
Most teams end up running both. Rules cover the must-haves (region, company size, vertical), and the predictive score ranks leads inside those limits.
How much data predictive lead scoring needs
Vendors publish different minimums, and they differ on whether you can see why a lead scored the way it did:
| Tool | Published minimum to train | Shows why a lead scored high? |
|---|---|---|
| HubSpot Likelihood to close | No minimum published | No. HubSpot calls the model “blackbox” machine learning (HubSpot) |
| HubSpot AI-built contact scores | 50 contacts: 25 converted and 25 not (HubSpot) | Yes. The AI recommends criteria you can see and edit |
| Salesforce Einstein Lead Scoring | 1,000 leads created and 120 converted in the last 200 days, per segment (Salesforce) | Yes. Shows the factors that contributed most |
| Microsoft Dynamics 365 Sales | 40 qualified and 40 disqualified leads in the training window (Microsoft) | Yes. Shows top influencing factors |
If you are below these numbers, start with rule-based scoring and build up a clean conversion history. Einstein can fall back to a global model built from anonymized data across Salesforce customers when you don’t segment your leads, but that model knows nothing specific about your buyers.
Predictive lead scoring in HubSpot
HubSpot now has three ways to score leads, and they are easy to confuse.
1. Likelihood to close and Contact priority (predictive). Available on Marketing Hub Enterprise and Sales Hub Enterprise (HubSpot). HubSpot sets both properties automatically, and you can’t edit them.
- Likelihood to close is the percentage chance a contact becomes a customer within 90 days.
- Contact priority groups contacts into Very High, High, Medium, and Low tiers based on that likelihood. Customers with a close date show as Closed Won. The tiers are relative, so the score range for each tier can shift over time.
- The model draws on web analytics and conversions, email engagement, CRM activity and properties, and company firmographics.
2. The lead scoring tool (rule-based). Available on Marketing Hub or Sales Hub Professional and Enterprise, under Marketing > Lead Scoring (HubSpot). You build fit scores from property values, engagement scores from actions, or combined scores from both. It replaced the legacy HubSpot score property, which stopped updating on August 31, 2025 (HubSpot Community). If an old workflow still filters on the legacy score, it is reading a frozen number.
3. AI help inside the lead scoring tool. AI Rules suggest scoring criteria from the last 90 days of activity, based on events that came before a lifecycle or deal stage conversion (HubSpot). On Marketing Hub Enterprise, you can also have AI build a contact score for you. Both leave you with rules you can read, which makes them the middle ground between hand-built scores and the closed Likelihood to close model.
How to use Likelihood to close in HubSpot
- Confirm the property exists. Open a contact record or the contact properties list and look for Likelihood to close and Contact priority. If they’re missing, the portal isn’t on an Enterprise Hub that includes them.
- Check it against reality first. Build a report of closed deals by the Contact priority each contact had. If Very High contacts don’t close more often than Low ones, fix your lifecycle stages before you trust the score.
- Build a segment for each tier. Filter contacts by Contact priority so SDRs have a working list for Very High and High.
- Trigger workflows on tier changes. Enroll contacts when Contact priority changes to Very High, then rotate an owner, create a task, or notify the rep.
- Keep a rule-based fit score next to it. A Very High contact from a country you don’t sell in is still a poor fit.
Limitations to know before you rely on the score
You can’t see inside HubSpot’s model. For Likelihood to close, HubSpot doesn’t expose weights or reasons. When a rep asks why one contact scored 87 percent and another 23 percent, nobody can answer from the data. That is fine for sorting a call list. It is a problem for disqualifying leads or reviewing a rep’s pipeline.
The model learns your past, not your plan. It is trained on what you closed before. If sales just moved upmarket or opened a new vertical, the model keeps favoring the old segment until enough new deals close.
Scores lag behind events. Models retrain on a schedule. For a lead who just requested a demo, the action itself is a stronger signal than the score.
A score doesn’t do anything by itself. A Likelihood to close of 87 percent doesn’t assign an owner, start an SLA clock, or tell a rep. Every action has to be built in a workflow that reads the score.
When rule-based or hybrid scoring is the better choice
- You are below the data minimum. Rules are the only option until there is enough history to train on.
- Your ICP is narrow and well understood. Explicit rules encode what you already know, and a model would mostly rediscover it.
- Your ICP is changing. Rules let you steer toward a new segment on the day you decide to. A predictive model catches up on its own retraining schedule.
The hybrid pattern works for most teams: rules for must-have fit criteria, with the predictive score ranking contacts that pass them.
Turning the score into action
Here is the part most predictive scoring guides skip. The score answers one question, how likely this lead is to close. Every question that follows is about routing:
- Who owns the lead, and are they at capacity?
- Where does the rep see it, and with what context?
- What happens when a contact jumps from Low to Very High overnight?
- Who reviews the leads in the middle tiers, where the model is least sure?
None of that lives in the scoring model. It lives between systems: the CRM that holds the score, the enrichment tool, the channel your reps work in, and the manager who decides the unclear cases. An orchestration layer is the category built for that gap. It sits above your systems of record, listens for changes like a score crossing a threshold, gathers the context, and brings a person in where judgment is needed. FlowRunner is built for that layer.
In practice, a FlowRunner flow connected to HubSpot can:
- Route by score and fit together. Very High plus ICP fit goes to an AE. Medium plus fit goes to an SDR for qualification. Low goes to nurture.
- Deliver the lead where reps work. Post to the owner in Slack with the score, source, recent activity, and a link to the record. The HubSpot to Slack routing guide shows the build.
- Ask a person about the middle band. FlowRunner uses human-in-the-loop, an execution pattern where AI agents pause autonomously, assemble the relevant context and the decision choices available, route to a human via their preferred channel, and resume the moment the human responds. A sales manager gets the borderline lead in Slack, email, WhatsApp, or by phone, decides, and the flow assigns it.
- Clean the inputs before they reach the model. Enrich and de-duplicate new contacts as they arrive, as in qualifying inbound HubSpot leads without manual research or enriching HubSpot contacts with Clearbit.
If you run Salesforce instead, the routing side works the same way; Salesforce lead assignment rules covers the native part and where it stops. To decide whether the routing work is worth automating at all, this guide to what’s worth automating treats it as a cost calculation.
Predictive lead scoring is one input to a qualified pipeline. Pair it with fit rules, check it against your closed deals, and build the routing around it, and the score earns a place in how your team works. If you’d rather have that routing built for you, our implementation team can set it up in your FlowRunner workspace.
Frequently asked questions
What is predictive lead scoring?
Predictive lead scoring uses a machine learning model trained on your past leads, both the ones that became customers and the ones that didn't, to estimate how likely each open lead is to convert. The model picks the signals and weights from your data instead of you assigning points by hand.
Does HubSpot have predictive lead scoring?
Yes. On Marketing Hub Enterprise and Sales Hub Enterprise, HubSpot sets a Likelihood to close property (the chance a contact becomes a customer within 90 days) and a Contact priority tier on each contact. Marketing Hub Enterprise can also build contact scores with AI in the lead scoring tool.
How much data do you need for predictive lead scoring?
It depends on the tool. Salesforce Einstein Lead Scoring asks for at least 1,000 leads created and 120 converted in the last 200 days, Dynamics 365 needs 40 qualified and 40 disqualified leads in the training window, and HubSpot's AI-built scores need at least 50 contacts, 25 converted and 25 not.
What is the difference between predictive and traditional lead scoring?
Traditional lead scoring adds points for rules your team writes, such as a job title or a pricing-page visit. Predictive scoring learns which signals mattered from your closed deals, so it can catch patterns nobody wrote down, but the weights are harder or impossible to inspect.