Forest management practices in recent times have evolved from focusing solely on timber production towards a more holistic approach incorporating ecological and social aspects. In order to support this movement, the Québec government has mandated and adopted a form of spatial organization termed COS (Spatial organization compartments). COS are subdivisions of the management unit designed to spatially aggregate a harvest area. The spatial aggregation of harvest areas ensures that the extent of the cut areas in the landscape is minimized, and it also reduces the amount of logging roads, thus reducing the complexity of forest operations. The delineation of COS is one of the first steps in management planning, and forest managers carry out this task manually. These units then become constraints for the calculation of Annual Allowable Cut (AAC). However, the boundaries obtained through manual delineation are rarely optimal. We propose a machine learning methodology combining the use of Graph Neural Networks (GNNs) and Reinforcement Learning to partly automate the process. Initially, the stand level data of a forest is encoded into a network of edges and graphs, which is then fed through the GNN to obtain topology aware graphs. After obtaining the dynamic graph that was passed through the GNN, the problem of boundary delineation can be formulated as a Reinforcement Learning problem. The model features an interactive process where the planning agent observes the graph's state, selects edges for optimal plans, receives rewards or penalties based on the guidelines set by the ministry, and the environment transitions based on the agent's actions. This machine learning approach has significant implications in the adaptation towards climate change. It becomes particularly invaluable during the replanning stage when natural disturbances impact planned harvest blocks, necessitating the swift generation of alternative optimal solutions.