A policy-driven method and an apparatus are disclosed. The method includes: obtaining at least one policy from a modeling function network element; receiving a first policy identifier triggered by a first model from a prediction function network element, where the first policy identifier is used to identify a first policy; determining the first policy in the at least one policy based on the first policy identifier, where the first policy includes action information corresponding to the first action; executing the first action in the first policy; and sending an execution result of the first action to a data service function network element or the prediction function network element, where the execution result of the first action is used as input data of a second model.