This article investigates the usage of Artificial Intelligence (AI) techniques in the prediction of network performance for Industrial Internet of Things (IIoT). In industrial environments, 5G Ultra Reliable Low Latency Communications (URLLC) are intended for serving critical services with very stringent latency requirements, such as those involving collaborative robots. Even if the flexible 5G New Radio (NR) design is able to achieve the target IIoT performances, the necessary spectrum resources need to be available and reserved for URLLC. A Quality of Service (QoS) prediction scheme is thus needed for anticipating performance degradation and undertake necessary actions, such as network resource provisioning or application adaptation, e.g. by entering an adapted mode. We explore the design of AI algorithms for QoS prediction in industrial environments, and compare different tools for regression and classification, including Neural Networks (NN) and K Nearest Neighbors (K-NN). We explore prediction based on Signal to Interference and Noise Ratio (SINR), or simply based on the position of robots within the plant. As the latency degradation event is rare in general, we observe that the training data is highly imbalanced leading to a low prediction accuracy. We show how the prediction performance can be enhanced by importance sampling techniques and by a modified detection threshold in what we call M-KNN scheme.