A propensity model is a statistical or machine learning model that estimates how likely each person in your database is to take a particular action in a set period: buy again, upgrade, respond to an offer, book an appointment or cancel. The output is a score or probability for every customer, which lets you decide who to contact, with what, and who to leave alone.
How a propensity model works
Every propensity model starts with a clear question tied to one action and one time frame, for example “Which customers will place another order in the next 60 days?” The model then learns from history. You take a past date, look at what you knew about each customer at that point, and record whether they went on to take the action. The model finds the patterns that separated the people who did from those who did not.
Typical inputs include:
- Purchase history: how recently, how often and how much someone has bought, and which categories.
- Engagement: email clicks, site visits, app logins, quote requests.
- Account details: plan type, tenure, region, how the customer was acquired.
- Service history: returns, complaints, failed payments.
Simple versions use weighted rules or logistic regression in a spreadsheet or BI tool. More advanced ones use gradient-boosted trees or similar methods in Python or a data warehouse. Many CRM, ecommerce and email platforms now include built-in scores such as “likely to purchase” or “predicted churn”. Churn prediction is one common use; lead scoring with a predictive model is another.
Why it matters
Most marketing budgets are spread evenly across people who will behave very differently. A propensity score lets you concentrate effort. A wine merchant might send its premium mixed case offer only to the top fifth of customers by purchase likelihood, instead of discounting to the whole list. A B2B software firm might have sales call the leads most likely to convert this month and leave the rest in a nurture sequence.
It also helps you avoid waste in the other direction. Customers who are very likely to buy anyway do not need a discount; offering them one gives away margin. The most useful targets are often the people in the middle, who might act with a nudge.
In the UK there is a legal side. Scoring people on their likely behaviour is profiling under UK GDPR. You need a lawful basis (often legitimate interests for marketing), and your privacy notice must tell people that you analyse their data this way and why. People have an absolute right to object to profiling used for direct marketing, so your systems need a way to honour that. Where a score leads to a decision with legal or similarly significant effects, such as refusing credit, stricter rules on automated decision-making apply; the Data (Use and Access) Act 2025 changed some of those rules, so check the ICO’s current guidance.
Common mistakes
- Data leakage. Training the model on information that was only known after the outcome (such as a refund issued because someone cancelled) makes it look accurate in testing and useless in practice.
- Targeting the sure things. The highest scores often belong to people who would buy without any marketing. Measure the lift, not just the response rate.
- No holdout group, so you cannot tell whether the campaign or the score did the work.
- Stale models. Customer behaviour shifts with seasons, prices and the economy; a model built two years ago may quietly stop working.
- Using sensitive data carelessly. Health, ethnicity or similar inferences fall into special category data and carry far stricter conditions.
How to act on it
Pick one decision you make repeatedly, such as who gets a win-back email or which leads sales calls first. Define the action and time window precisely. Before building anything, check whether your platform already offers a predictive score and how it performs on your data: compare the scores it gave last quarter with what customers actually did.
Use the scores to create a few bands, assign an action to each, and hold back a random group in each band to measure the real effect. Update your privacy notice before you start, and record your legitimate interests assessment. Deciding where scoring is worth the effort, and how it fits a wider retention and acquisition plan, is part of my digital marketing strategy and consulting. For acquisition, the related idea is a value-based lookalike audience, which asks an ad platform to find new people who resemble your most valuable customers.