A lead scoring model ranks leads by two things: how well they fit your ideal customer (fit) and how much they have engaged (engagement). Build it from closed-won data, decay engagement points over time, and route leads above the threshold within an hour. Predictive lead scoring comes later, once your CRM clears the vendor's data minimums.
Most scoring models fail for a social reason. Marketing builds the model, sales sees a few bad leads labeled hot, and reps stop looking at the score. The fix is a model sales helped design, tested against real outcomes, with a clear promise about what happens when a lead crosses the line. Here is how to build that model in order, and how to tell whether your CRM is ready for AI scoring.
A lead scoring model combines fit and engagement
Keep the two dimensions separate. HubSpot's scoring tool defines fit scores as qualifying records by demographic properties such as job title, company size or annual revenue, and engagement scores as qualifying them by actions such as website visits, newsletter sign-ups and CTA clicks (HubSpot). A combined score then pairs them, with labels from A1 to C3, where the letter is fit and the number is engagement.
The split matters because each dimension answers a different question. Fit asks whether you should want this account. Engagement asks whether the person is active now. A C1 lead (poor fit, very active) can be a student, a job seeker or a competitor, while an A3 lead (great fit, quiet) belongs in a nurture or ABM target account list. Collapsing both into one number hides that difference from the rep.
Set lead scoring criteria from closed-won deals
Lead scoring criteria should come from your own wins and losses. Pull the last 12 to 24 months of closed-won and closed-lost opportunities, then compare the firmographics and behaviors that preceded each. Weight the attributes that separate the two groups, and drop the ones that look important in theory but appear equally in both.
Research on CRM data supports this. A 2025 peer-reviewed study in Frontiers in Artificial Intelligence tested fifteen classification algorithms on a company's real CRM lead data and found that features such as lead source and lead status improved conversion prediction, with gradient boosting performing best (Frontiers in AI). Behavior deserves extra weight now. Gartner found that 67 percent of B2B buyers prefer a rep-free experience and 45 percent used AI during a recent purchase (Gartner), so much of the buying journey happens in self-serve signals before anyone talks to sales.
Score decay keeps a B2B lead scoring model honest
Engagement goes stale, so engagement points should expire. A webinar attended in March says little about intent in October. HubSpot lets teams set score decay intervals of 1, 3, 6 or 12 months (HubSpot), and three months is a sensible starting point for sales cycles under a year. Fit points rarely need decay, since company size and industry change slowly.
Negative scoring belongs in the same pass. Subtract points for personal email domains, careers-page visits, unsubscribes and job titles you never sell to. Then audit the data feeding the model: a score built on a misspelled industry field or duplicate contacts will misroute leads no matter how well it is weighted. Our guide to CRM data quality lists the fields worth fixing first.
Predictive lead scoring needs more data than most teams have
Predictive lead scoring is only as good as its training data, and each CRM sets a floor. HubSpot's AI scores need at least 50 contacts, 25 converted and 25 not, and are part of Marketing Hub Enterprise (HubSpot). Microsoft Dynamics 365 Sales needs at least 40 qualified and 40 disqualified leads closed within a training window of three months to two years (Microsoft Learn). Salesforce's Einstein implementation guide, last updated in 2023, lists at least 1,000 leads created in the last 200 days with at least 120 converted, and when all leads are scored together without enough conversion data, Einstein falls back to a global model built on anonymized data from many customers (Salesforce).
Clearing the minimum is the easy part. In Salesforce's State of Sales 2026 survey of 4,050 sales professionals, 74 percent said they were focusing on data cleansing and 51 percent of leaders using AI said disconnected systems were slowing their AI initiatives (Salesforce, vendor survey). Run a rules-based model for a quarter, fix the data it exposes, then switch on predictive scoring and compare the two side by side.
Lead scoring only pays off with a fast routing SLA
A score threshold is a promise from marketing to sales, and speed is what makes it worth keeping. A Harvard Business Review study of online leads found that firms contacting a lead within an hour were nearly seven times as likely to qualify it as firms that first tried even an hour later (Harvard Business Review, 2011). A hot score that sits in a queue overnight is a cold lead by morning.
Write the SLA next to the model. Define who owns a lead above the threshold, how fast they respond, how many attempts they make, and what disposition they record. Review those dispositions monthly and adjust weights where sales keeps rejecting leads. If an AI agent handles first touch, the same rules apply, which we covered in AI SDR vs human SDR.
Feed lead scoring stages back into Google Ads and Meta
A working score can also train your ad platforms. Google Ads lets advertisers specify the value they want to maximize, such as "sales revenue, profit margins, or lead scores," and recommends bidding on a funnel stage with at least 15 monthly conversions (Google Ads Help). Meta's Conversion Leads goal, which works with Instant Form lead ads, fits advertisers generating at least 200 leads a month whose target stage converts at between 1 and 40 percent within 28 days (Meta for Developers).
Send the stage your score predicts best, usually a sales-accepted lead, and let the platform optimize toward it. The mechanics of the Google side are in our post on Google Ads lead quality. If you want a RevOps review of your scoring model and routing, talk to our team.
