A method for data processing includes obtaining from multiple different organizations (24, 52, 54, 56) customer relationship management (CRM) records (110) and communication records (90). Reference vectors (130) of feature values are computed for the communication records in a training set. Global weights are computed for the feature values by evaluating the reference vectors for all of the different organizations. For each organization, respective company weights are computed by evaluating specifically the reference vectors computed over the CRM records and communication records belonging to the organization. For each person belonging to a given organization, respective user weights are computed for the feature values by evaluating specifically the reference vectors computed over the communication records that identify the person as the user. The weights are applied in order to assign the communication records that are not in the training set to respective ones of the CRM records.