The Internet of Things can be an effective way to manage the demand side and perform demand response in thermal grids. The concept provides new models for predicting demand and gathering data in a smart and inexpensive fashion. This research investigates the deployment of room-specific demand response in a district-heated office building in Southern Finland. It is fulfilled by controlling set-point temperatures on thermostatic radiator valves. By developing a predictive algorithm on a cloud platform and gathering feedback from the environment and the users, the paper presents findings of load shifting on room level and about local discomfort. The results indicate that heating power can be increased and decreased at the thermostat, even by 104% and 47% respectively. On average, this accounts for 40 W/m(2). However, the larger the set-point temperature variation is, the larger local discomfort grows. Indoor air temperature readings do not give enough evidence of thermal comfort. Hence, demand response with room-level accuracy should be deployed in areas of the building in which flexibility is least noticed. (C) 2019 Elsevier Ltd. All rights reserved.