Correlation is not causation - causal inference gives ML engineers tools to answer "would changing X cause Y?". Potential Outcomes Framework: for each unit, define Y(1) (outcome if treated) and Y(0) (outcome if untreated). Average Treatment Effect (ATE) = E[Y(1) - Y(0)]. We can never observe both, so we need assumptions. Randomized A/B tests are gold standard but expensive and sometimes unethical. Observational methods: 1) Propensity Score Matching: match treated and control units with similar probability of treatment given covariates, reducing selection bias. 2) Instrumental Variables: find variable Z that affects treatment D but only affects outcome Y through D - exogenous variation for causal ID. 3) Difference-in-Differences: compare pre/post changes between treatment and control groups, assuming parallel trends. 4) Double Machine Learning (DML): use ML to partial out confounders from both treatment and outcome before estimating causal effect - achieves sqrt(n) convergence even with high-dimensional confounders. Uplift Modeling: estimate individual-level treatment effects for targeting interventions (marketing emails, discounts) to those who will respond positively. Libraries: DoWhy (Microsoft) for causal graph modeling, CausalML (Uber) for uplift modeling, EconML (Microsoft) for heterogeneous treatment effects.