High resolution aerial photos combined with accurate map data represents a perfect data set for training artificial intelligence models. The ‘KartAi’ project is an innovation project in public sector aimed at developing Ai-methods that detects buildings not in the cadastre or the building map dataset. Thereafter involving the property owner/citizen in a digital dialog and validate or crowdsource more detailed data. The foundation for this is high quality datasets for training and validating the different Ai-models. High resolution aerial photos are collected in large parts of Norway on a regular basis – often yearly – in a collaboration between federal and municipal. Thereby there exists a vast amount of extremely detailed image data combined with building map data and cadastre data. However, training the Ai-models have uncovered that minor errors and ‘skewed’ photos and/or vector data affects the results of the segmentation of roof tops/buildings. Therefore the KartAi projects has made fine tuned and accurate training data sets in several geographical areas optimized for training on detecting and segmenting buildings. In several large scale experiments, a multitude of existing models, newer models and own models have been training and validated. Additionally we have included LIDAR-height data to enhance the precision of segmenting between the likes of roofs and terraces. Training the models on the existing data yields good results. However, when finetuning with the high accurate data – the models show impressing results. Spatial Ai projects like KartAi are at the mercy of volumes of good training data. Our experience show that even more accurate data sets improve the models even further. Therefore, the project has made efforts that have resulted in the release of the training data sets publicly – as well as all of the results data for the different models and approaches that have been developed. This is an effort into developing a more open living lab for Spatial Ai in Norway. Our hope is that sharing the knowledge and data created can ensure that other Ai-models have easier access to high resolution and high accuracy data – to train models in the open living lab – and apply the models internationally where data is scarcer.