Beneficence

Explanation

Beneficence checks that an AI system actively does good for the customer and stakeholders — providing helpful, accurate, suitable guidance — rather than merely avoiding harm. In banking this matters because a chat or RAG assistant should genuinely improve a customer's financial outcomes (e.g. flagging a cheaper repayment option or a relevant grievance channel) rather than giving hollow or self-serving answers. It is measured by the Beneficence Score, which rates whether outputs deliver real, relevant benefit to the user. To implement it, instrument the response layer of summarisation/chat and RAG pipelines: evaluate sampled and high-stakes responses for helpfulness, correctness, and suitability using an automated rubric-based scorer supported by human spot-checks, and log the score, the response, and the supporting evidence (retrieved sources, reasoning) for audit. Scoring can run inline for critical advice and in batch for ongoing monitoring against quality dashboards. The threshold requires the beneficence score to stay above a defined floor; if a response falls below that floor it triggers a breach — the answer is improved or escalated to a human, the gap is investigated, and the event is logged, aligning with RBI Ethics and Autonomy and NIST GOVERN 5.1 stakeholder-benefit objectives.