Knowledge
Related service:
Rikard Jonsson
Rikard Jonsson is Founder & CEO of Hey Sid and a five-time entrepreneur with a background in B2B SaaS, sales, and brand building. He believes B2B marketing is overcomplicated and writes about going back to basics: visibility, positioning, and consistent presence among the accounts that matter.
Propensity Modeling in B2B: Predicting Which Buyers Are Ready, Not Just Which Accounts
TL;DR
Propensity modeling in B2B predicts the probability a specific buyer takes an action inside a defined window, not just whether an account looks active.
Account scoring and intent data answer "which company?" A person-level buyer propensity score answers "which individual, and how ready?"
A usable model needs four things: a defined outcome, person-level features, clean training data, and an action layer that routes the score to a play.
Most models fail on operationalisation, not math: a score nobody acts on changes nothing.
The payoff: engage the right individual the moment readiness climbs, instead of spraying a whole account.
Related reading: Account Scoring Models That Align Sales and Marketing | Intent Data for B2B: Find In-Market Accounts | Person-Based Marketing: The Complete B2B Guide
Your intent platform lights up an account. Sales calls in. The person who picks up has no budget, no timeline, and no idea why you reached out. The account was in-market. The buyer was not ready.
This is the gap propensity modeling in B2B is built to close. Account scoring and intent data tell you which companies show activity. A person-level propensity model tells you which individual buyer is likely to act, and how soon.
Most teams still model at the account level because that is where the data is easiest to buy. But B2B purchases are made by people: a buying group of typically 6 to 10 stakeholders, each with a different role, timeline, and reason to engage. Score the account and you know where to knock. Score the person and you know whose door to knock on, and when.
This guide covers how to build a B2B propensity model that predicts buyer readiness at the individual level: the outcome to predict, the features that matter, the modeling approach, and the action layer that turns a score into pipeline. It is written for RevOps, demand gen, and marketing ops teams who already run account scoring or intent data and want to get closer to the person.
What Is Propensity Modeling in B2B, and Why Account Scores Miss the Buyer
A propensity model outputs a probability that a specific entity takes a defined action inside a time window. The output is usually expressed from 0 to 1, or scaled to 0 to 100. A propensity to buy model, for example, might estimate a 0.18 chance that a given contact books a meeting in the next 90 days.
In most B2B stacks, that entity is an account. The problem: the entity that signs a contract is a person, and an account can be in-market while the specific person you are reaching is not the buyer, or is not ready.
Three approaches get confused here. They answer different questions.
Approach | Question it answers | Unit | Typical output |
|---|---|---|---|
Account scoring | Which companies fit and look engaged? | Account | Fit and engagement score |
Intent data | Which companies are researching now? | Account | In-market flag or surge |
Person-level propensity | Which individual is likely to act, and when? | Person | Probability of an action in a window |
Account scoring ranks companies by fit and aggregate engagement, and it is the right tool for territory planning and prioritisation. Intent data flags accounts researching a category before they fill in a form. Both operate at the account level. Neither tells you that the VP of Marketing who joined your webinar and revisited pricing twice is far more likely to convert than the analyst who downloaded one report.
That distinction matters because readiness is not evenly spread across a market. At any given moment, most of your addressable buyers are not in an active buying cycle. Modeling readiness at the person level is how you find the small share who are, and reach them before a competitor does.
Propensity modeling does not replace account scoring or intent data. It sits on top of them, using account-level signals as features and adding the individual behaviour that account scores average away.
How to Build a B2B Propensity Model: Step by Step
You do not need a data science team of ten to start. A first useful model can run on the data already in your CRM, marketing automation platform, and enrichment tools. The sequence below moves from definition to a score sales can act on.
Step 1: Define the Outcome and the Time Window
A model is only as clear as the event it predicts. Pick one concrete, measurable outcome: a booked meeting, an opportunity created, a demo attended. Then fix a window, commonly 30, 60, or 90 days.
"Likely to buy" is too vague to train on. "Likely to create an opportunity within 90 days" is a label you can attach to historical records. The outcome and window define your positive class for every step that follows.
Step 2: Assemble Person-Level Features
Features are the inputs the model learns from. For a buyer propensity score, group them into three types.
Fit attributes: role, seniority, function, and the firmographics of their account (industry, size, region). A CFO and a junior analyst at the same account carry different propensity.
Behavioural signals: email engagement, ad interactions, content downloads, event attendance, repeat visits to high-intent pages such as pricing. These are the raw material behind most buying signals that predict pipeline.
Contextual signals: account-level intent surges, existing relationships, and product usage where relevant. Account intent data becomes one feature among several, not the whole verdict.
The step that separates a person-level model from an account model is contact-level behaviour: individual engagement history, not the account's summed activity.
Step 3: Build a Clean Training Set From Wins and Losses
Supervised models learn from labelled history. Pull records of contacts who did and did not hit your defined outcome, then label each row positive or negative.
Two issues sink most training sets. The first is class imbalance: your positive class is often under 5% of records, so the model can score 95% "accuracy" by predicting "no" every time. Handle it with resampling or class weighting. The second is data leakage: never include a feature that only exists after the outcome, such as "opportunity amount," or the model will look brilliant in testing and useless in production.
Step 4: Start Simple, Then Add Power
Begin with logistic regression. It is fast, interpretable, and gives every feature a readable weight, which helps you sanity-check what the model thinks drives readiness. It is a stronger starting point than the hand-tuned point systems most predictive lead scoring in B2B began as.
When a baseline works, move to gradient boosting, such as XGBoost or LightGBM. Tree-based models capture non-linear interactions, for example seniority mattering only when paired with recent pricing-page visits. Complexity is a second step, not a first.
Step 5: Validate, Calibrate, and Check for Leakage
Test on data the model has never seen. A time-based holdout, training on older records and testing on newer ones, mirrors production better than a random split. Use AUC to rank models and precision and recall to judge them at the threshold you will actually deploy.
Then calibrate. A propensity score is only trustworthy if a "0.30" means roughly 30% of those contacts convert. Uncalibrated scores mislead the sales team that acts on them. Re-check for leakage whenever a feature looks suspiciously predictive.
Step 6: Operationalise the Score Into a Play
A score in a notebook changes nothing. Route it into the systems where revenue teams work, and attach an action to each tier.
Translate the probability into bands, for example high, medium, and watch, and define what happens at each. A rising high-propensity contact might trigger coordinated advertising and a sales touch on the same individual. This is where a person-level model pays off: you are not activating a whole account, you are engaging the specific buyer the model flagged.
Step 7: Monitor Drift and Retrain
Buyer behaviour shifts, and a model trained on last year's patterns decays. Track live performance against predictions and watch for drift in your input features. A quarterly retrain is a reasonable default for most B2B teams, with a faster cadence if your market or motion changes quickly. Treat the model as a system you maintain, not a project you finish.
Where B2B Propensity Models Break
Most failures are not statistical. They are decisions made before or after the modeling.
Modeling the Account When the Person Decides
An account-level score averages a buying group of 6 to 10 people into a single number. That is useful for prioritisation and misleading for outreach: it cannot tell you which stakeholder to contact. If your goal is to reach a buyer, model the buyer.
Training on Leaky Features
Any feature that only exists after the outcome inflates test performance and collapses in production. Opportunity value, close date, and post-sale fields are common offenders. Audit every feature for whether it would have been available at prediction time.
Ignoring Class Imbalance
When positives are rare, a naive model learns to predict "not ready" for everyone and reports high accuracy. Weight the classes or resample, and judge the model on precision and recall for the positive class, not overall accuracy.
Treating the Score as Static
A propensity to buy model is a living asset. Without monitoring and retraining, its accuracy erodes quietly while the team keeps trusting it. Schedule retraining before performance, not after complaints.
Building a Score With No Action Layer
The most common failure is organisational. A well-built model produces clean scores that sit in a dashboard because no play is attached. Decide what a high score triggers before you build it, or the model never touches pipeline.
Tools You'll Need to Build and Act on Propensity Scores
A propensity program has two halves: the stack that produces the score, and the stack that acts on it.
For the model itself, you need a data foundation (your CRM and marketing automation as the source of truth, a warehouse such as BigQuery or Snowflake if volumes are large, and enrichment to fill gaps), plus a modeling layer (Python with scikit-learn or a no-code predictive tool). Several platforms also ship prediction out of the box.
6sense is one of the strongest predictive platforms on the market, scoring accounts on likelihood to buy across a large intent network. But 6sense predicts at the account level. It tells you an account has entered a buying stage, not which individual on the committee is ready to act.
Demandbase offers one of the most complete enterprise ABM suites, with account intelligence and native advertising in one place. Its prediction and targeting are account-level, largely firmographic and IP-based, so the output is the in-market account rather than the ready person.
ZoomInfo gives you deep contact and company data, which is valuable raw material for features. But it is a data source, not a readiness model: enrichment is an input to a propensity score, not the score itself.
Contact-level ad platforms such as Influ2 show that person-level targeting works in practice. The open question for most teams is the action layer: once a specific buyer's propensity climbs, can you reach that individual across more than one channel, consistently?
That activation gap is where Hey Sid fits. Hey Sid runs person-level advertising (Always On), automated LinkedIn outreach (Precision Connect), and done-for-you thought leadership (Authority Builder) against the same named individuals, coordinated as The Influence Loop. A propensity model tells you which buyer is ready. Hey Sid is designed so that buyer then meets a coordinated presence built around them, not a single cold touch.
Risk Ident attributed 2.5x shorter sales cycles and 40% higher engagement to that coordinated model (client-reported).
Explore how it works: heysid.com/how-it-works
Conclusion: Score the Buyer, Then Act on the Score
Account scoring and intent data earned their place: they tell you which companies deserve attention. Propensity modeling in B2B goes one layer deeper, predicting which individual buyer is likely to act and when, so your team spends effort on the person closest to a decision. The math is the easy part. The value comes from a clear outcome, honest features, and an action layer that turns a score into a coordinated play against the right person.
If you want the activation side handled while your model does the predicting, that is the problem person-based marketing is built for, and it is why most buyers stay invisible until they are ready to act.
Book a demo: heysid.com/demo
FAQ
What is propensity modeling in B2B?
Propensity modeling in B2B uses historical data to predict the probability that a specific account or contact takes a defined action, such as booking a meeting or creating an opportunity, within a set time window. The output is a score you can rank and act on. Person-level propensity focuses that prediction on the individual buyer rather than the whole company.
What is the difference between propensity modeling and intent data?
Intent data detects when an account is researching a category, usually from third-party signals across the web. Propensity modeling predicts the likelihood of a future action and can operate at the individual level. Intent is one input; propensity is the prediction. Teams increasingly feed account intent into a person-level model as a feature rather than treating it as the answer.
How is a buyer propensity score different from lead scoring?
Traditional lead scoring assigns fixed points to attributes and actions, tuned by hand. A buyer propensity score is produced by a statistical model trained on real wins and losses, so it weights features by their actual predictive value and outputs a calibrated probability. Predictive lead scoring in B2B is the machine-learning evolution of the point-based approach.
How much data do I need to build a B2B propensity model?
There is no fixed minimum, but you need enough labelled positive outcomes for the model to learn a pattern, not just a handful. Because positive classes are often under 5% of records, teams with thin history usually start with a simple logistic regression and clear features before attempting more complex models. More history and cleaner labels matter more than model sophistication.
Should I model at the account level or the person level?
Use account-level propensity for territory planning and prioritising which companies to work. Use person-level propensity when the goal is outreach, because a purchase is decided by a buying group of typically 6 to 10 people and you need to know which individual to engage. Mature programs often run both, with the account score setting priority and the person score directing action.
How often should a propensity model be retrained?
A quarterly retrain is a reasonable default for most B2B teams, with a faster cadence if your market, product, or go-to-market motion shifts quickly. The trigger is drift: when live performance or the distribution of input features moves away from the training data, retrain. Monitoring should be cotinuous even if retraining is periodic.
Sources
Related: Account Scoring Models That Align Sales and Marketing | Intent Data for B2B: Find In-Market Accounts | Why 95% of Your B2B Customers Are Not Ready to Buy Yet
