<p>Organisations accumulate years of operational decisions made by experienced professionals. Embedded in this historical data is a valuable asset: a representation of how the organisation actually makes decisions.</p> <p>This presentation explores how artificial intelligence (AI) can turn that accumulated experience into models capable of supporting - and, under controlled conditions, autonomously executing - operational decisions.</p> <p>Starting from a real insurance claims use case involving more than half a million historical cases, Gabriel De Dominicis examines how AI models can learn recurring decision patterns from previous human behaviour and identify situations in which a decision can be safely automated. The approach combines deterministic eligibility rules, machine learning models, explainability, and explicit risk controls, rather than relying on an unconstrained black-box prediction model.</p> <p>The initial results demonstrated a high level of precision for selected automated decisions, even with a deliberately restricted set of features, while indicating the potential to free capacity equivalent to thousands of working days each year. The objective, however, is not simply to maximise predictive accuracy.</p> <p>False positives and false negatives have different operational and economic consequences. For this reason, the optimisation target becomes the value generated by automation within an acceptable level of risk. The project addresses this through an Automation Gain framework that evaluates correct automated decisions, decision costs, and the potential consequences of errors together.</p> <p>The presentation will discuss the practical lessons emerging from this experience: how to select appropriate historical information, prevent data leakage, combine business constraints with learned behaviour, explain individual AI decisions, monitor model performance over time, and progressively move from AI prediction to controlled autonomous execution.</p> <p>Ultimately, the goal is for an organisation to learn from thousands of its previous decisions and turn that collective experience into a continuously improving operational capability, rather than simply reproducing individual human decisions.</p>