Your marketing team builds a churn prevention campaign. You pull a list of "at-risk" customers from the CRM, send a retention offer, and wait. Two weeks later, half the customers on the list had already churned before the email arrived. The other half were never going to churn at all. The list was based on gut feel and a single filter: "hasn't logged in for 30 days." That is not predictive scoring. That is a guess wearing a spreadsheet.
Propensity modeling replaces guesswork with probability. Instead of filtering on a single field, a propensity model analyzes dozens of signals across a customer's history and assigns a score: how likely is this specific person to take a specific action? The action can be purchasing, churning, converting from trial to paid, or engaging with a campaign. The score turns a binary yes/no segment into a ranked list, so your team prioritizes the customers most likely to respond.
This article is part of our series on AI-powered marketing intelligence, which covers how prediction systems depend on connected customer data. Where the pillar explains recommendation systems broadly and our AI decisioning satellite covers automated action selection, this article goes deep on the scoring and prediction mechanics that feed those systems. A propensity score is often the input that recommendation engines and decisioning platforms use to determine what to show, send, or offer each customer.
What propensity modeling is and how propensity scores work
Propensity modeling is a statistical technique that predicts the probability of a customer taking a specific action. The output is a propensity score: a number (typically 0 to 100) that represents how likely that customer is to buy, churn, convert, or engage within a defined time window.
The mechanics are simpler than most guides suggest. A propensity model takes historical data about customers who already took the target action and customers who didn't. It identifies which attributes and behaviors differ between the two groups. Then it applies those patterns to your current customers to generate a score for each one.
For example, a churn propensity model might analyze 500 customers who churned in the last year and 2,000 who didn't. It finds patterns: churned customers averaged 2 logins per month (vs. 12 for retained), filed 3+ support tickets in their last 60 days, and hadn't upgraded their plan in 6 months. Now, when it sees a current customer with 3 logins last month, 4 recent support tickets, and no plan changes, it assigns a high churn propensity score.
The propensity score itself is not a prediction that something will happen. It is a ranking tool. A customer with a score of 85 is not "85% likely to churn." They are significantly more likely to churn than a customer scoring 30. The value is in relative ranking: your retention team works the highest-scored accounts first, and your marketing team tailors messages by score band.
Types of propensity models: purchase, churn, conversion, and engagement
Four propensity model types cover the majority of marketing use cases. Each predicts a different action and requires different input signals.
Model type | Predicts | Key input signals | Typical use case |
|---|---|---|---|
Purchase propensity | Likelihood of buying a product or upgrading | Browse history, cart activity, past purchases, billing tier | Cross-sell campaigns, product recommendations |
Churn propensity | Likelihood of canceling or not renewing | Login frequency, support tickets, feature usage decline, payment failures | Retention campaigns, success team prioritization |
Conversion propensity | Likelihood of converting from trial/free to paid | Feature adoption, time-in-trial, team invites, billing page visits | Trial nurture sequences, sales outreach timing |
Engagement propensity | Likelihood of responding to a specific campaign | Email open history, channel preference, recency of last interaction, content affinity | Send-time optimization, channel selection |
Purchase propensity is the most established type. E-commerce teams have used it for decades: score customers by their likelihood to buy, then target high-propensity customers with product recommendations and offers. The customer scoring model uses purchase recency, frequency, and monetary value (RFM) as a baseline, then layers on browsing behavior, wishlist activity, and category affinity.
Churn propensity is the highest-impact model for SaaS teams. Identifying at-risk accounts before they cancel gives your retention team a window to intervene. The signals that predict churn are often obvious in hindsight but invisible when they live in different tools: login frequency drops (product database), support tickets spike (helpdesk), payment method fails (billing tool), and email engagement disappears (marketing platform). No single tool sees all four signals.
Conversion propensity helps product-led growth teams decide which trial users to nurture and when. A trial user who invited two teammates, connected an integration, and visited the pricing page twice has a different conversion propensity than one who logged in once and never returned. Predictive customer scoring based on trial behavior lets your team focus outreach on users most likely to convert.
Engagement propensity determines who to contact and how. Not every customer responds to email. Not every customer responds on the same day. Propensity scoring for engagement predicts which channel, time, and content type each customer is most likely to respond to. This is where propensity models feed directly into AI decisioning systems that automate per-customer send decisions.
Propensity modeling without a data science team: practical approaches
Most guides assume you have a data science team, a Python notebook, and a warehouse full of labeled training data. That describes maybe 5% of companies under 200 people. The other 95% need predictive scoring that works with the tools and team they already have.
Approach 1: Rule-based scoring in your CRM. HubSpot, Salesforce, and Attio all support custom scoring properties. You define the rules: +10 points for logging in this week, +20 for being on a paid plan, -15 for filing a support ticket, +25 for visiting the pricing page. The total is a predictive score. It is not statistically rigorous, but a rule-based customer scoring model built on 5-8 high-signal fields outperforms complex ML models built on incomplete data.
Approach 2: RFM segmentation. Recency, Frequency, Monetary value. Score each customer on three dimensions: how recently they took action, how often they take action, and how much revenue they generate. Combine the three into a composite score. This is a simple but effective propensity scoring technique that e-commerce teams have used since the 1990s. It works for any business with repeat transactions.
Approach 3: Spreadsheet regression. Export your customer data, label which customers took the target action, and run a logistic regression in a spreadsheet or a free tool like Google Colab. This takes an afternoon, not a quarter. The output is a set of weights for each input field, which you can implement as a scoring formula in your CRM.
All three approaches share one prerequisite: the input fields must exist in a single system. If login frequency lives in your product database, support ticket count lives in Zendesk, and billing status lives in Stripe, none of these approaches work until that data reaches your CRM or scoring tool.
Why propensity models fail when customer data is fragmented
The most common failure mode is not bad math. It is missing inputs.
Consider a churn propensity model that trains on CRM data. The CRM has contact details, lifecycle stage, deal value, and email engagement. That is four dimensions. But the signals that actually predict churn live elsewhere:
Login frequency: product database
Feature adoption: product database
Support ticket volume: helpdesk (Zendesk, Intercom)
Payment failures: billing tool (Stripe)
NPS response: survey tool
A propensity score built on CRM data alone is like predicting the weather using only temperature. You have one dimension when you need five. The model might catch customers who stop opening emails, but it completely misses the customer who logs in every day, opens every email, but just had three failed payments and filed two angry support tickets.
This gap is exactly what enterprise CDPs and data platforms claim to solve. They centralize customer data in a warehouse, build ML features, train models, and push scores back to your tools. The approach works for companies with data engineering teams and six-figure budgets. For everyone else, it is a 6-month project that delays predictive scoring indefinitely.
The practical alternative: get the data to move between tools directly. If your CRM can see login frequency, support ticket count, billing status, and feature adoption alongside its native email engagement data, even a simple rule-based scoring model becomes surprisingly accurate. The propensity score improves not because the math got better, but because the inputs got complete.
Building predictive signals from connected tools
The fastest path to working propensity models is not hiring a data scientist or buying an ML platform. It is connecting the tools that hold the signals your model needs.
Step 1: List the signals that predict the target action. For churn: login frequency, support tickets, payment status, feature usage, email engagement. For conversion: feature adoption, team invites, billing page visits, time in trial. Write down the tool that holds each signal.
Step 2: Sync those signals into your scoring tool. For most teams, the scoring tool is the CRM. Sync login count from your product database, open ticket count from your helpdesk, subscription status from your billing tool, and feature flags from your application. Each signal becomes a contact property in the CRM.
Step 3: Build the scoring rules. Use the CRM's native scoring feature. Assign point values to each signal based on what you've observed in churned or converted customers. Start simple: 5-8 fields, round-number weights. You can refine later.
Step 4: Act on the scores. Route high churn-score contacts to your success team. Trigger a retention email sequence for medium scores. Exclude low-score contacts from discount campaigns (they don't need a discount to stay). For conversion scoring, trigger sales outreach when a trial user crosses a threshold.
This is the approach we take at Oneprofile. We sync data between your tools so the fields your propensity model needs are available where you build and act on scores. Stripe billing status, Zendesk ticket counts, product database login frequency, and marketing engagement data all flow into your CRM on a schedule you control. Each field syncs with precision: only changed values are updated, so your scoring properties always reflect the latest customer state.
The result: your CRM has the 5-8 fields that predict churn, conversion, or purchase intent. Your propensity scoring rules work on complete data instead of a single dimension. Your team acts on ranked lists instead of gut-feel filters. And you built it in an afternoon, not a quarter.
No warehouse. No ML pipeline. No data science team. Just connected tools and a scoring formula that reflects what your data already knows about your customers.
What is propensity modeling in marketing?
Propensity modeling uses historical customer data to predict future behavior. It assigns each customer a propensity score indicating how likely they are to take a specific action: purchase, churn, convert, or engage.
Do you need a data scientist to build a propensity model?
Not for basic models. Simple propensity signals like days since last login, support tickets filed, and billing status predict behavior well. CRM-native scoring tools can use these fields without any ML pipeline.
What data does a propensity scoring model need?
At minimum: behavioral data (logins, feature usage), transactional data (purchases, billing status), and engagement data (email opens, support tickets). The more complete the customer profile, the better the score.
How is propensity scoring different from lead scoring?
Lead scoring ranks prospects by fit and engagement. Propensity scoring predicts a specific outcome for any customer, not just leads. A churn propensity model scores existing customers, not prospects.
Why do propensity models produce inaccurate predictions?
Usually because the model trains on incomplete data. If billing status, support history, and product usage live in separate tools that don't share data, the model sees a fraction of each customer's behavior.