Smart cities provide services through IoT sensor/actuator networks. Today, more than ever, energy efficiency is a must, particularly in building automation scenarios. RFID, cell phones or tablets could be used, but they are devices to be carried or must be logged in. Cameras have privacy issues. Activity recognition and occupancy are tasks that can be accomplished with device-free WiFi sensing to estimate thermal load. Humans in a room „interfere“ with Tx-Rx signals, producing CSI (Channel state Information) „images“ that can be recognized by Artificial Neural Networks. With CSI we have immediate change detected and not after a temperature change is detected by the control loop. Enabling energy saving by anticipative thermal comfort control strategy (estimated 25% energy saving). Artificial Neural Networks are powerful learning tools, but what happens if the environment changes? That means, we are looking for AI that adapts easily (FewShot Learning). Using ESP32 modules CSI „images“ can be used to count people and identify human activities. The objective is to learn in one environment and then apply FewShot to different furniture arrangements.