Pain assessment in horses represents a significant challenge in veterinary medicine, since these animals are unable to verbally express their discomfort, and their pain signals are often subtle and difficult to interpret. Traditional assessment methods, which rely on behavioral observation, can result in inaccurate and variable diagnoses, compromising adequate pain management. The main objective of this work was to develop an artificial intelligence-based tool to identify pain signals in horses, using the Facial Action Units (FAU) described in the Horse Grimace Scale (HGS). For this, a set with 1673 images was obtained, annotated into two classes in Roboflow: “Pain + ” and “Pain − ”. The YOLOv8n model was trained, and the model demonstrated a good balance between precision and recall, being effective in detecting the “Pain −” and “Pain + ” classes. The precision, recall, and mAP metrics stabilized at high values (approximately between 0.98 and 0.99) after about 50 epochs, indicating that the model achieved near-optimal performance. On the test data, the model achieved a 97% accuracy rate, demonstrating high accuracy in classifying between the two categories. The creation of this automated system represents a significant advance in the field of animal welfare and the detection of pain signals. The method stands out for its non-invasive and non-disturbing nature, ensuring the safety and quality of life of the animals.