How a propensity model works
Three things must be fixed before any modelling begins: the action, the population and the time window. Will this contact purchase, within the next month, given what is known today? Change any of the three and you have a different model. The output is a probability for each individual, usually turned into ranked groups, because the ranking is more useful than the raw number.
The model learns from history. Rows of past customers, each described by what was known about them at a point in time and by what they went on to do, are given to an algorithm that finds which combinations of attributes preceded the action. Applied to current contacts, it produces scores. Accuracy is checked by holding back part of the history, scoring it as though the outcome were unknown, and seeing whether the high-scoring group really did act more often than the low-scoring one.
Why propensity models matter
They let you spend differently on different people. A retention offer sent to everybody is expensive and wasted on customers who were never going to leave; sent to the group a churn prediction flags as at risk, the same budget goes further. The same logic decides who receives a discount, who gets a phone call, and which audience an ad campaign is built from.
They also make forecasting less superstitious. A pipeline weighted by modelled probability is a more honest view of next month than a list of deals somebody felt good about.
Where propensity models go wrong
The most damaging error is treating a probability as a fact about an individual. A high score does not mean this person will buy. It means people who look like this have bought more often. Decisions should be made at the group level, where probabilities behave, not by declaring one customer certain.
Then there is leakage, where a field that only becomes true after the outcome is included in the model. Something filled in by sales when a deal closes will make the model look excellent in testing and useless in practice. Related to this is the self-confirming loop: if only high-scoring contacts are ever approached, the results agree with the model whether or not it was right.
Finally, models inherit the gaps in your tracking. If consent choices or blocked scripts mean whole segments of behaviour are missing, the model learns from the visitors you happened to observe.
How to act on it
Begin with the decision you want to make, then build the smallest model that informs it. Compare its ranking against an obvious baseline, such as recency of purchase, order value or recent engagement, and keep the model only if it beats that. Many businesses find the simple rule was close enough, which is a useful and cheap thing to learn.
Always hold back a control group treated as though the model did not exist. Without one you can show that the targeted group acted, but not that the targeting caused it. And make sure the measurement underneath is sound before modelling on top of it, because analytics and tracking that misses events will produce a confident model of the wrong thing.
