AI lead scoring: what it is, how it works and how to apply it in your company

Lead scoring is the method for ranking your leads by their probability of becoming customers, so that sales calls first whoever is closest to buying. For years it has been done with hand-assigned points: +10 for the job title, +5 for opening an email, +20 for visiting the pricing page. AI lead scoring replaces that points table with a model that learns from your past conversions and calculates, for every lead, a real probability of purchase that changes as the lead moves.

In this article you'll see why the points model stops working as soon as volume grows, how a predictive model works, what data it needs and how to put it in place without a data science team.

What lead scoring is

A lead is any person or company that has shown interest and whose contact details you hold: they filled in a form, subscribed, asked for information, bought once. Lead scoring gives each one a value that sums up how much they deserve your time right now.

The goal isn't the score itself but the decision it enables: who the sales team calls today, who gets an offer, who is left to mature and who is never contacted again. Good lead scoring cuts the time sales wastes on contacts who won't buy and stops a hot lead going cold while it waits its turn.

Traditional lead scoring: why the points model falls short

Classic lead scoring combines two kinds of criteria: profile (job title, industry, company size, region) and activity (opens, clicks, visits, downloads). Each criterion adds or subtracts points and, above a threshold, the lead goes to sales.

It works while volume is small and the team knows its customers. It stops working for four reasons:

• The weights are opinions. Nobody knows whether opening three emails is worth more or less than visiting the pricing page; it gets decided in a meeting and never touched again.
• It doesn't learn. If the market shifts or a new channel comes in, the table keeps scoring like it did on day one.
• It rewards the ideal profile and misses the real buyer. A junior researching for their director scores low and gets dropped; the director never visited the site.
• It's blind before the form. All the anonymous browsing before someone identifies themselves doesn't count, and that's usually where the intent is.

The usual outcome: a list of "hot leads" that sales learns to ignore because they don't convert any better than the rest.

How AI lead scoring works

A predictive lead scoring model doesn't start from rules but from your history: which leads ended up buying and which didn't, and what each group did before the decision. From that history it learns which combinations of signals anticipate a purchase and applies what it has learned to every new lead.

In practice the model works in three steps:

1. It unifies the signals. Web and app browsing (including before identification), emails, forms, webinars, social, calls and, if they're already a customer, purchases and tickets. All on a single person or company profile.
2. It calculates the probability. For each lead, the model estimates the probability of converting within a given window, say the next 30 days. A lead who attends a webinar, downloads a whitepaper and visits pricing in the same afternoon moves up immediately; it doesn't wait for the weekly review.
3. It corrects itself with the outcome. As leads convert or don't, the model records it and adjusts which signals weigh more. Over time it reflects your market, not an assumption from two years ago.

What data you need and what metrics to ask for

A predictive model is only as good as the history that feeds it. You need three things:

• Enough past conversions. A few hundred leads with a known outcome (bought / didn't buy) is a reasonable starting point; with fewer, the model learns noise.
• Signals tied to the right person. If the form, the email and the browsing live in three different tools that don't talk to each other, the model only sees part of each lead. This is where a Customer Data Platform comes in: it resolves identity and delivers the full history.
• Accuracy metrics, not promises. Ask the vendor how the model is measured: a ROC curve or an AUC above 0.8, and the comparison between the conversion rate of the top-scoring decile and the average. If the best-scored 10% converts three times more than the average, the model is doing its job.

Lead scoring in B2B and B2C

The term comes from B2B, where every lead is an opportunity worked by a person. But the same logic applies in B2C under a different name: purchase propensity. A retailer with stores and ecommerce doesn't score "leads"; it scores its customers and subscribers to decide who gets the campaign, who gets the offer and who gets nothing. The models are the same; what changes is the volume and the activation channel.

And in companies with both kinds of customer, scoring needs to work at two levels: the person browsing and the company they belong to. FLYDE resolves both identities on the same data, so an account's score picks up the signals of all its contacts.

How FLYDE does it: Lead2Customer

Lead2Customer is FLYDE's predictive lead scoring model. It trains on your conversion history, scores every lead with a conversion probability that updates with each interaction, and shows its accuracy metrics inside the platform so you know how far to trust it.

Because FLYDE is a CDP, the model works on the unified profile: anonymous browsing is attached once the user identifies themselves, and the later purchase is used to retrain. The score flows out to your CRM, your email tool and your campaigns, and from FLYDE Brain you can ask in plain language "which leads above 60% haven't been contacted this week?" and get the list.

How to put it in place in five steps

1. Define the conversion. What counts as success: a meeting, a purchase, a contract? The model predicts exactly what you teach it.
2. Gather the history. Export the leads from the last 12–24 months with their outcome and their interactions. If they sit in three tools, this is the step that takes longest and is worth most.
3. Train and measure. Let the model learn, then compare its prediction with what happened in a period it hasn't seen. If the top decile doesn't clearly convert more than the average, review the data before going further.
4. Activate with thresholds. Decide what happens in each band: above 60%, a call within 24 hours; 30–60%, an email sequence; below 30%, nurturing. Keep the thresholds simple and reviewable.
5. Close the loop. Feed the outcome of every contacted lead back to the model. Without this step the model stops learning and turns back into a points table with a nicer look.

Frequently asked questions

How many leads do I need for AI lead scoring to work? As a guide, a few hundred leads with a known outcome. Below that volume, a simple, well-maintained points model may be the better option until you build up history.

Does predictive lead scoring replace the CRM? No. The CRM is still where sales works each lead; the model tells them which one to focus on first and why. Typically the score appears as one more field on the CRM record.

What's the difference between lead scoring and purchase propensity? They're the same kind of model applied to different audiences: leads in B2B, customers and subscribers in B2C. In both cases the output is a conversion probability per person.

What you gain

Sales stops chasing leads who won't buy and reaches the ones who will before they go cold. Marketing stops arguing about weights in a spreadsheet and measures the model with a number. And the "hot leads" list goes back to being a list someone actually wants to call.

If you'd like to see how many of your current leads have a high probability of converting, request a demo and we'll calculate it on a sample of your data. And if you're not sure your data is ready for a model like this, the self-assessment tells you in two minutes.