Description

In a rapidly evolving Global AI landscape, large language models (LLMs) and autonomous agents are increasingly shaping critical decision-making processes—from fraud detection and credit scoring to welfare distribution. Unlike deterministic systems, these models operate on probabilities and confidence scores rather than certainties. This raises a fundamental challenge: how can we guarantee fairness, accountability, and trust when AI outcomes are inherently uncertain? This session explores how to embed Responsible AI principles into the very architecture of probabilistic AI systems. Moving beyond pure prediction, we demonstrate how to design solutions that can explain, justify, and remain fully auditable—drawing on live implementations in financial oversight and public service delivery. Through real-world case studies, we will showcase: a) Credit risk assessment: An AI agent that not only provides eligibility scores, but also explains its confidence intervals using real DIA data. b) Welfare benefit allocation: LLM-driven recommendations that include rational, human-readable visualisations for both auditors and citizens. c) Fraud detection: A monitoring tool that highlights suspicious transactions and provides a transparent reasoning trail for auditors before decisions are made.