Although importance of animal welfare assessment is recognized, practical evaluation methods in zoos have not been sufficiently discussed. This paper presents a series of practical steps to improve animal welfare using a case of a Japanese black bear at the Kyoto City Zoo, and examines the advantages and challenges of three animal welfare evaluation methods (risk assessment of animal welfare and behavior assessment by humans and AI) used in the process. The first approach was through human ratings using a welfare risk assessment sheet. This assessment sheet comprised of questions on the environment, husbandry routines and state of the animal. The zoo staff (N = 23; including keepers, vets and researchers) input the scores into the assessment sheet before and after implementing several environmental enrichment (May 2019 and March 2020). The second and third approaches were based on 960 hours behavioural observations by humans and behavioural estimation by AI, respectively. These behaviours were recorded every 5 min by 2 observers through an infrared camera attached at the ceiling of the outdoor enclosure between September 2019 and February 2020. The same recorded videos were also analysed using feature values extracted from video frames with deep learning techniques. The results revealed that all the methodologies detected changes in the welfare states of the animal. However, variations in the strength and limitations existed among the methodologies. Human ratings using the assessment sheet detected overall improvements in environment and behaviours before and after implementing enrichment. However, variations in the scores among the evaluators existed. Behavioural observations by humans and behavioural estimations by AI produced similar results. The concordance rate between the human behavioural observations and AI observations was approximately 77%. Besides, by incorporating information about the spatial location of the animal into the AI observations, the rate improved to approximately 84%, the value close to the concordance between 2 human observers (approximately 85%). Furthermore, the behavioural changes before and after implementing an enrichment item detected by both human and AI observations were similar. However, the accuracy of AI behavioural estimation fluctuated across behaviours. These results together suggest that selecting and combining appropriate assessment methodologies in each purpose based on the strength of each approach are important.