Predictive lead scoring trains a model on the leads you won and lost, then scores each open lead by how much it looks like the wins. Every CRM that sells one agrees on that much. Ask how many closed deals the model needs, though, and the answers stop agreeing: Microsoft asks for 40 qualified and 40 disqualified leads, Salesforce for 1,000 leads with 120 conversions.
That gap is the question behind the term, and it decides whether a predictive model is worth switching on at all. If the model types are new to you, what a lead scoring model is, with the three types compared sets rules-based, predictive and AI scoring side by side. This page goes one level down: how the model learns, how much history it needs, and three ways it quietly goes wrong.
If you already suspect your history is too thin, score your accounts against your ICP for free while you read, and check the reasons against what your reps know.
What is predictive lead scoring?
A points table is a guess you write down. You decide a pricing page visit is worth 5 and a VP title is worth 15, and the CRM adds them up.
The predictive version replaces the guess with arithmetic. It looks at every lead that reached a decision, won or lost, and works out which field values separated the two groups. The output is usually a number from 0 to 100, read as how closely an open lead resembles your past wins. Some products call it a likelihood or a conversion probability, which is the same idea with a different label.
Three things have to hold for it to work:
Enough decided leads. The model learns from outcomes, so a lead still in progress teaches it nothing.
Fields that were filled at the time. It can only weigh what the CRM held when each lead was decided.
A next customer who resembles the last ones. It ranks open leads by similarity to old wins.
Most of this page is about what happens when one of those three fails.
How a predictive lead scoring model learns its weights
Under the hood, most predictive scoring runs one of two model families. Neither is exotic.
Logistic regression gives every input field one weight. Say leads at companies with 50 to 200 employees close more often than leads at companies under 10 people. The headcount field gets a positive weight, and the model adds up the weighted fields and squeezes the sum into a probability. It reads almost like a points table, except the points came from your data.
Gradient-boosted trees build hundreds of small decision trees, each one correcting the mistakes of the trees before it. They catch combinations a single weight misses, such as a director title that matters only above a certain company size. That makes them better at ranking. It also makes them much harder to explain, which comes back later.
Whichever family it is, a lead scoring model built with machine learning goes through the same loop:
Pull every lead with a final outcome inside the training window.
Turn each field into numbers the model can read. An industry becomes a set of yes or no flags, one per industry.
Fit the weights that best separate won from lost.
Score every open lead, then retrain on a schedule as new outcomes arrive.
A side note on the accuracy figures that show up in sales decks. Say 5 in 100 of your leads close. A model that predicts "lost" for every lead is then 95% accurate and tells you nothing. The better question is how many of the top-scored tenth went on to close, compared with everyone else, because that is what a rep's call list actually feels.
How much history predictive lead scoring needs
Here is where the help pages stop agreeing. Each CRM that ships a predictive or AI-built score states its own minimum in its help docs. Microsoft's figure below was read in its Dynamics 365 predictive lead scoring docs on Nov 18, 2026, HubSpot's in its knowledge base on Oct 30, 2026, and Salesforce's in Salesforce Help, as our Salesforce AI lead scoring guide quotes it.
Approach | What it learns from | History it needs before it scores |
|---|---|---|
Microsoft Dynamics 365 Sales predictive lead scoring | Leads qualified or disqualified in a window you pick, from 3 months to 2 years | 40 qualified and 40 disqualified leads, created and closed in that window |
Salesforce Einstein Lead Scoring | Leads that converted against Leads that did not | 1,000 Leads created in the last 200 days, 120 of them converted; with all Leads scored together, a global model from many orgs' anonymized data stands in below that |
HubSpot score built with AI (Marketing Hub Enterprise) | Contacts who made a lifecycle stage change you pick | 25 contacts that converted and 25 that did not |
A regression you train yourself | Your own CRM export of won and lost leads | About 10 wins for every field the model weighs |
A points table or an ICP fit score | Criteria you write down | None |
The spread looks odd until you notice what a model runs short of first. It is rarely leads. It is wins, the rarer of the two outcomes.
Statisticians have a rough guide for this called the one in ten rule: a regression needs about ten examples of the rarer outcome for every input it weighs. By that rule, 40 wins supports a model that weighs about four fields, and 120 supports about twelve. A team with 30 closed-won deals and a CRM full of fields has enough history for maybe three of them, which is roughly the size of a points table you could have written by hand.
So the honest answer to "how much history" is a count of wins divided by the fields you want weighed. Leads still in progress do not count, however many there are.
Where predictive scoring goes wrong
Say the history clears the bar. Three problems still survive it, and none of them shows up on a setup screen.
It learns from fields that were empty. Reps fill in industry and headcount on the leads they work, usually after a first call. Leads nobody called stay blank, and most of them end up lost. The model sees "industry is filled in" sitting beside the wins and learns that a filled field predicts a win. A new inbound lead arrives blank, so it scores low until a rep touches it, and the score ends up tracking rep attention rather than buyer fit. Modelers call this leakage.
It ranks more of who you already attract. A model trained on last year's wins ranks open leads by how much they look like last year's buyers. That works while the funnel is healthy. Suppose most of your wins came from agencies and this year you are selling to software companies: the new segment has almost no wins in the training data, so by the one in ten rule the model cannot learn anything about it, and those leads score low by default. The HubSpot AI lead scoring guide makes the same point about HubSpot's own scores.
It hides why. A rep sees 64 and has to decide whether to call. Some products do better here than others: Dynamics 365 lists the top influencing factors behind each lead's score. Even then, a factor like "lead source is webinar" says what correlated with winning, not whether this company fits what you sell, and a rep who cannot check a score in ten seconds goes back to calling whoever signed up last.
None of this makes predictive scoring a bad idea. It makes it a late one.
CRM lead scoring before you have the history
Anyway. Most teams reading this have a few dozen wins, a CRM with half its firmographic fields empty, and a quarter's worth of leads to rank. Two moves fit that situation, and both still help on the day you do switch a predictive model on.
Fill the fields first. Enrich every company in the CRM, won, lost and open alike, so headcount, industry and technology stack are filled across the board rather than only where a rep happened to type them. In Oneprofile, Enrich company from domain costs up to 15 credits a company and Find technology stack up to 5.8. By default, both check their data sources in an order set by accuracy, not price, and a company lookup that finds nothing costs nothing. Then pull in what your own systems know: plan and payment status from Stripe, product usage from PostHog, ticket history from your support desk. A future model trained on those records learns from facts instead of from which leads a rep opened.
Score fit against a written ICP. Ask a question that needs no history: which companies match the customer you describe? The ICP scorer reads the ideal customer profile and buyer personas you write in Business context and rates each company on its own, never against the rest of the list. You get a band (strong at 70 or above, possible from 40 to 69, poor under 40), a 0 to 100 score, and the two or three facts that drove it. A company with too little data gets no band and a list of what was missing, never a poor score, which is the opposite of the leakage problem above. It costs 3 credits a company. ICP scoring with AI covers how to write the profile from your closed-won accounts.
Fit is half the picture, and it is worth being plain about which half. A fit score says nothing about timing:
"The ones showing a reason right now: hiring for the role your product replaces, or posting about the problem. Fit tells you who could buy. A signal tells you who might buy this month."
Utku Zihnioglu, CEO of Oneprofile
Write the score and its reason back. CRM lead scoring only changes behavior if reps see it where they work. Create a number field for the score and a text field for the reasons in your CRM first, because the export maps onto fields the CRM already has, then export band, score and reasons to them. With Skip empty values on, a company that came back with no score leaves the existing value alone. The setup screens differ by CRM, so the Salesforce guide and the HubSpot guide above walk through each one.
On cost: scoring alone is 3 credits a company, so the Free plan's 200 credits a month score 66 companies that already carry their firmographics, with up to 3 integrations, 3 tables and 100 rows per table. With Enrich company from domain added, a company costs up to 18 credits. Team starts at $20 a month for 1,000 credits.
Then keep counting wins. When you have about ten per field you want weighed, and those fields were filled before each deal closed, predictive lead scoring has what it needs, and the fit score you wrote can be one of the fields it learns from. Until then, score your accounts against your ICP free and give every rep a reason they can check.

