Customer Relationship Management is an every day task for companies, even the ones dealing with Small Data. We are more interested here by Lead Scoring that refers to the practice of calculating and assigning a score to leads (business contacts or qualified prospects) of the company. In this paper, we present one way of building a Lead scoring model with a Bayesian network using a small amount of data. In addition to its ability of handling uncertainty, Bayesian networks are knowledge representation models that can be built from expert knowledge. In our specific context, we then propose to build our Lead scoring model from expertise and apply usual heuristics to decrease the complexity of our model (parent divorcing, NoisyOr). We specifically propose three ways of estimating the parameters of our NoisyOr submodels. The only data available is used to validate our approach, with good precision and recall results on a small set of 23 examples.