The horses' responses to exercise are commonly monitored using physiological variables, nonetheless physical and mental states can also be expressed through body language. The aims of this study were: (i) to identify how facial expressions and other behavioural variables change in ranch horses after a routine workday, and (ii) to investigate if these changes can be used as indicators of physical tiredness by relating them to known variables of physical fitness and workload. Fourteen crossbred ranch horses were assessed pre- and post-workday on two farms, recording the body language, physiological and workload variables. Statistical analysis consisted of four stages: (i) comparisons between the sampling times (pre- vs post-workday) using linear mixed-effects models with repeated measures and a paired Wilcoxon test; (ii) selection of the most powerful variables by applying Kaiser-Meyer-Olkin test and principal components analyses (PCA); (iii) evaluations of the relationships within these selected variables utilizing PCA and Spearman rank coefficients; and (iv) identifying a critical level of the most robust behavioural indicators using a non-hierarchical cluster analysis. Results showed that after a workday the horses increased the frequency/duration of body language indicative of resting. They also decreased the frequency/duration of body language indicative of attention and movements to avoid flies. However, some of these behaviours are also shown when horses are in pain, leading us to suggest that some ranch horses were probably experiencing a combination of of tiredness and slight soreness. Of particular interest, because of the ease with which it can be assessed on the farm and generalized to other situations, we suggest that the frequency of shifting weight between the forelegs has potential to be used as an indicator of physical tiredness in horses. The results can also be used in the development of tools to improve the welfare of ranch horses as well as horses used in other activities, although more research is needed to validate this assumption.