FROM AUTHORISATION TO AGENCY: REGULATING MACHINE-EXERCISED PAYMENT AUTHORITY IN INDIA

Author

Akshat Nigam , Arushi Rajagopala, Lavya Bhasin, Yohann Titus

Abstract

 Agentic payment systems alter a familiar assumption in digital payments: that the person whose credentials authenticate a payment is also the person who has decided to make that particular payment. India’s payment architecture already permits substantial forms of prior and delegated authority. UPI Circle permits full delegation to another user within defined limits, UPI Reserve Pay allows a customer to block funds for multiple future debits to a merchant. The legal difficulty therefore is not that Indian law requires a human to approve every digital payment.

The more precise difficulty is that agentic AI can be given authority to decide the transaction itself. The customer may authenticate a delegation, while the agent later selects the merchant, product, timing and transaction amount within a mandate. This separates four questions that existing payment regulation tends to connect - whether the customer was authenticated, whether the mandate was validly authorized, whether the agent was authorised to act, and whether the individual transaction fell within that authority. A fifth question follows - who bears the loss when an authorised agent makes a disputed or out-of-scope decision?

This article examines these questions through India’s current authentication framework, e-mandates, UPI Reserve Pay, UPI Circle and the 2026 customer-liability amendments. It argues that existing mechanisms can accommodate bounded forms of agentic payment, but do not yet provide a sufficiently explicit legal architecture for machine-exercised payment authority. It proposes a three-tier framework distinguishing automated execution, bounded agentic authority and open-ended agentic authority, supported by verifiable mandates, agent identity, transaction-level decision records, evidentiary rules and liability allocation according to the source of failure. The objective is not to create a separate body of law for every AI payment, but to make explicit the legal significance of the machine’s decision between the customer’s delegation and the resulting payment.

Keywords

Agentic AI, digital payments, UPI, authentication, delegated authority, e-mandates, UPI Reserve Pay, UPI Circle, payment liability, machine-exercised payment authority

Full Text:

Download Paper PDF

References

India does not need to begin the legal analysis of agentic payments from the proposition that digital payments require fresh human authentication for every transaction. That proposition is already disproved by the country’s own payment architecture. E-mandates permit subsequent recurring payments without additional-factor authentication within the applicable framework. UPI Reserve Pay permits multiple debits against a customer-created reserve. UPI Circle permits delegated payment authority, and NPCI has extended that concept to devices and software profiles.

The legal novelty of agentic payments lies elsewhere. An AI agent can be given authority not merely to execute a payment already selected by the customer, but to decide what the payment should be. That changes the relationship between authentication and authorisation. The customer can authenticate a delegation, the agent can make a decision, the payment system can authenticate or process the resulting instruction, and the customer can still dispute whether that particular decision was within the authority granted.

The central analytical distinction is therefore between authentication of the customer, authorisation of the mandate, authorisation of the agent, authorisation of the individual transaction, attribution of the agent’s decision, and allocation of liability. Existing law contains pieces of each question, but does not yet express them as a single framework for machine-exercised payment authority.

The emerging industry response is already moving in that direction. P3P demonstrates a live attempt to combine mandate-based payment rails with agent identity, delegated authorisation and auditable execution. Mastercard has demonstrated authenticated agentic commerce.

The appropriate regulatory response is consequently not a blanket prohibition on autonomous payment. Nor is it a blanket extension of existing authentication exemptions. A better approach is a three-tier framework that treats automated execution, bounded agentic authority and open-ended agentic authority differently. Bounded agentic payments should be capable of operating without fresh transaction-level human approval where the delegation is authenticated, the agent is identifiable, the authority is objectively bounded, and the system can produce a reliable record showing compliance with the mandate. Open-ended authority should require stronger regulatory safeguards.

The principal contribution of this approach is to shift the legal focus from whether a human clicked “pay” to whether the system can establish a legally meaningful chain from human delegation to machine decision to payment. India already knows how to authorise a payment in advance. The unresolved question is how to authenticate, attribute and allocate responsibility for a decision that the customer authorised a machine to make, and this question can only be answered with time, as technology progresses, and the legal framework follows.

Share your valuable work from Social Media Buttons