Most of the cost of developing software systems occurs after its initial release, thus challenging software engineers to embrace change. This key aspect of software is known as the quality attribute modifiability. With the adoption of AI components, change to software is not only happening in source code but also in AI-models and the data that the model is using. In the proposed research, quality requirements for changing AI-models, data and code are identified, new practices are defined, integrated into a reference architecture, and validated with an industry field study on mature AI-based systems.