Before We Give AI Agents the Keys to Our Money: Rethinking Trust, Intent and Accountability in Agentic Finance

Abstract Agentic finance changes a foundational assumption of digital payments: that a human is directly involved in initiating or approving each consequential transaction. As artificial-intelligence agents begin to perform tasks on behalf of users, conventional controls built around authentication, device trust and transaction authorisation may no longer be sufficient.This paper develops a conceptual framework for trusted agentic finance centred on a fourth requirement: intent integrity. It argues that financial systems must distinguish among what an agent is permitted to know, what it can infer, and what it is permitted to do. The paper proposes bounded autonomy based on distinct agent identity, purpose-bound authority, separation of information access from transaction authority, transaction and contextual limits, enhanced verification for anomalous actions, traceability and ex-ante liability arrangements. It also considers low-and-slow compromise, third-party concentration and systemic risks arising from common AI models, infrastructure and agent frameworks.The paper argues that India's digital public infrastructure gives the country an opportunity to embed these controls into agentic payments at the infrastructure layer rather than retrofit them after adoption.