Modern machine learning excels at pattern recognition but often fails to support decision-making, as it cannot distinguish correlation from causation. This is a critical limitation in high-stakes domains, where relying on statistical associations can reproduce historical biases embedded in the data. To address this, we apply do-calculus within a causal Bayesian network (CBN) framework to estimate the effect of residential energy-efficiency interventions (specifically, external wall insulation) on household gas consumption. By encoding structural assumptions in a directed acyclic graph, we derive post-intervention distributions from observational data, disentangling causal identification from statistical inference. This enables estimation of both average and subgroup-specific treatment effects, revealing substantial behavioural heterogeneity: households under high energy burden show significantly smaller energy savings post-intervention. Ultimately, this work illustrates how causal ML can address the biases and limitations of predictive models, and how formal tools like do-calculus can transform ML systems into more robust instruments for policy and decision-making under uncertainty.