Non-Maleficence

Explanation

Non-maleficence checks the ethical 'do no harm' principle directly: across the automated recommendations an AI system makes, what fraction avoid causing a harmful outcome to a customer or third party — for example a recommendation that pushes an unsuitable product, denies someone unfairly, or exposes them to financial detriment. It matters because in BFSI automated advice and decisions touch people's livelihoods, and a system that is accurate on average can still inflict concentrated harm that regulators and ethics frameworks expect institutions to actively measure and minimize. It is measured by the Harm Avoidance Score, computed as Harm Avoidance = 1 - (Count(Harmful Outcomes) / Count(Total Decisions)) for automated recommendations: one minus the proportion of decisions that led to a harmful outcome. Implement it by defining what counts as a harmful outcome with risk and compliance, instrumenting recommendation logs so each decision can later be classified as harmful or not (via complaints, redress cases, outcome reviews, or human adjudication), and aggregating the score with the supporting case evidence over RAG and Summarization and Chat flows. The threshold is a Harm Avoidance Score of 0.995, meaning fewer than 0.5% harmful outcomes; falling below it is a breach that triggers investigation into the harmful cases and remediation before the system continues issuing recommendations.