In smart environments, there is an increasing demand for scalable and autonomous management systems. In this regard, energy efficiency hands out challenging aspects, for both home and business usages. Scalability in energy management systems is particularly difficult in those industry sector where power consumption of branches located in remote areas need to be monitored. Being autonomous requires that behavioural rules are automatically extracted from consumption data and applied to the system. Best practices for the specific energy configuration should be devised to achieve optimal energy efficiency. Best practices should also be revised and applied without human intervention against topology changes. In this paper, the Internet of Things paradigm and machine learning techniques are exploited to (1) define a novel system architecture for centralised energy efficiency in distributed sub-networks of electric appliances, (2) extract behavioural rules, identify best practices and detect device types. A system architecture tailored for autonomous energy efficiency has interesting applications in smart industry—where energy managers may effortlessly monitor and optimally setup a large number of sparse divisions—and smart home—where impaired people may avoid energy waste through an autonomous system that can be employed by the users as a delegate for decision making.