Artificial intelligence (AI)-enabled networking technologies have emerged in radio network research. As data is the key driver of these AI-enabled technologies, RAN (Radio Access Network) architectures are evolving towards openness, interoperability and reconfigurability to support RAN data collection and control. However, different AI-enabled approaches consume large amounts of data of different types, formats, and sizes. Previous works, prototypes, and testbeds have not gone into detail about managing the provision of significant amounts of data and analyzing the impact of data collection on RAN performance. In this paper, we therefore propose a unified framework for RAN data collection and provisioning. Based on the framework, we implement a data collection and provision (DCP) demo, a plug-in network function (NF) for data collection on user-/ data-plane of RANs in a cloud base station (BS). Two potential use cases of DCP NF are demonstrated. We evaluate the performance of DCP NF and its impact on RANs in terms of the delay to provision the collected data, and the computational and network bandwidth overhead. Our results show that DCP NF enables RAN data collection and provisioning to facilitate RAN intelligence.