In order to increase manufacturing performance in terms of resource and efficiency, it is required to approach the process physical boundaries manually by empirical trial and error, mostly driven by experience of few experts. Die-Sinking (DS) EDM is a very complex nonlinear time-varying process, which requires a control sequence of parameters called policy to optimize metrics like machining time, electrode wear and surface quality. Current research aims to use novel Reinforcement Learning (RL) to synthesize policies for model-free control problems like in DS-EDM. The optimal control policy is obtained by training the system for each application case with an heuristic sub-optimal policy. An IoT stack architecture will then be introduced as possible way to deploy the abovementioned AI solution, where a centralized collection of such control policies for different use-cases can be leveraged with Fleet Learning (FL) to share the knowledge across workshop machines.