This paper presents a multiagent architecture and algorithms for collaborative learning in distributed and heterogeneous business systems, where the participating agents have local, incomplete knowledge used to make predictions about parameters of a business transaction.We propose two collaborative learning strategies which differ in the nature and amount of information that is exchanged during collaboration, and which are hence suitable for different organisational settings. The first algorithm relies on the exchange of information about a transaction instance, whereas the second algorithm uses qualitative information provided by individual agents, such as the results of predictions from the agent?s local perspective. We apply the architecture and strategies to a distributed supply chain prediction problem. Experiments run on a large real-world order data set indicate that our approach effectively improves the learning performance based on limited additional communication between the participating agents.