In this paper, we describe an AI-based system that recognizes the activity status of several people from video streams during brainstorming meetings. Deep learning is often used to recognize video characteristics but requires a huge amount of computer resources. This makes it difficult to keep track of the activities of multiple people whose circumstances change. On the other hand, many trained models of one person’s motion recognition have been developed and are available. We propose to use the existing technology but to be able to do that we need to identify a single person’s activities within a group context. This is achieved by segmenting the video and cropping the area with a person, identifying the activity using pre-existing trained models. The activity of the group is recognized by a production rule system based on individual activities. To achieve our goal, we introduce the concept of atomic action to describe activities and propose categories of atomic actions. High-level collaborative categories that indicate the status of a group during collaborative meetings are based on the CIAO model. This paper ends with the results of the first experiments we conducted using video recordings of actual students’ work sessions.