Until now, initiating a payment has had a fairly clear protagonist: a person.
They select a product, access a channel, authenticate, and authorize a transaction. Behind that experience sits an infrastructure designed to recognize who is acting and under what conditions they are allowed to do so.
Artificial intelligence is beginning to introduce an important variation: the individual may still make the decision, but no longer execute the transaction directly.
The sequence is starting to change. And the transformation is not simply about who interacts with the merchant. It is about whether the infrastructure can determine that an agent is actually acting within the conditions established by the user.
The User Can Delegate, but Cannot Disappear
Mastercard is already testing infrastructure in Chile that allows AI agents to execute transactions under conditions defined by the user. In September 2026, the company announced that six banks in the country had completed real transactions using infrastructure designed for agentic payments.
Visa is also developing infrastructure for this model and has expanded its Agentic Ready program into Latin America. Its approach includes mechanisms to identify legitimate agents, enable them to make payments, and verify that they are acting with the user’s approval.
These developments highlight an important distinction: delegating an action does not mean authorizing every action an agent may take.
If a user specifies that an agent may purchase certain products up to a defined amount, for example, authorization should remain tied to those conditions. The payments infrastructure must therefore interpret more than the credentials used to initiate the transaction.
Authorization Begins Before the Payment
In a traditional payment flow, authentication and authorization are closely linked to the user’s direct action.
With AI agents, authorization can be granted in advance and subject to specific conditions: who may act, for how long, for what types of purchases, within what spending limit, and under which circumstances.
The International Monetary Fund identifies precisely this transition as one of the significant changes agentic AI could bring to payment systems. Its analysis distinguishes three layers: intent, authorization, and settlement, and notes that agents may affect not only transaction initiation, but also authorization, compliance, settlement, and operational resilience.
This changes the question the infrastructure must answer. It is no longer enough to know who is paying. It will also be necessary to determine who is acting, on whose behalf, and within what scope of authorization.
Identifying the Agent Is Not Enough
India offers a particularly relevant indication of where this challenge may be heading.
The National Payments Corporation of India (NPCI) is developing a registry to verify and monitor AI agents conducting transactions through UPI, the Unified Payments Interface, as part of a protocol for agentic payments. The initial model is expected to focus on frequent, low-value transactions, with more complex instructions potentially added later.
Registering the agent addresses one part of the problem: identifying which software is acting within the system.
But another question remains: was that specific transaction actually within the scope authorized by the user?
This introduces new elements that the infrastructure will need to manage: agent identity, spending limits, delegation validity, purchase conditions, traceability of the instruction, and the ability to stop or reject a transaction that falls outside those parameters.
Knowing who the agent is, by itself, does not resolve transaction authorization.
More Actors Also Mean More Risk Signals
An AI agent can operate at a speed and scale that a person would find difficult to sustain. It can evaluate multiple options, initiate several transactions, and respond automatically to merchant conditions.
That introduces another consideration: not all automated activity will necessarily be legitimate.
The infrastructure will need to distinguish between an agent correctly executing a delegated instruction and behavior that shows signs of fraud or misuse. Decisions therefore cannot rely solely on a credential or on the agent’s identity. Behavioral patterns and transactional context will also need to be assessed.
This is where artificial intelligence appears again from a different angle: not only as the actor initiating transactions, but also as a technology used to identify anomalous behavior.
The New Flow Will Also Need to Be Protected and Tested
If a transaction carries information related to identity, authorization, limits, and the conditions of delegation, protecting the data and cryptographic keys involved becomes even more critical.
At the same time, this new payment flow will need to be tested before it reaches production.
What happens if the agent interprets an instruction differently than intended? What if it exceeds a spending limit? What if it attempts a purchase outside the agreed conditions? What if it receives an unexpected response from the merchant? What if an authorization expires while the transaction is still in progress?
These are scenarios that should not be encountered for the first time in production.
Infrastructure Ready for AI Agents
The emergence of AI agents requires payments infrastructure to answer several questions at once: Who is acting? What is the agent authorized to do? How should the transaction be processed? And which signals indicate that behavior has moved outside the expected parameters?
In this context, CLAI PAYMENTS® Technologies provides capabilities that can address each of these requirements within a unified architecture.
AZ7® orchestrates transactions and applies the rules that determine how each operation should be processed. CRYPTGRID™ protects the cryptographic keys and data that underpin transaction security. AXIA™ uses artificial intelligence and machine learning to identify patterns associated with potentially fraudulent behavior. PAYTESTER® reproduces transactions from physical and digital channels to validate different scenarios before they are moved into production.
Together, these capabilities address agentic payments across several critical dimensions of the operation: deciding how transactions should be processed, protecting the information that sustains trust, identifying risky behavior, and validating that the payment flow performs as intended.
The next transformation in payments may not be a new payment method, but a new actor capable of initiating one.
And when that actor is artificial intelligence, the infrastructure will need to understand more than what is being paid for. It will need to know who is acting, on whose behalf, under what conditions, and with what level of trust.
If your organization is preparing for this next stage in payments and is evaluating the technology required to support AI-agent payments within its architecture, complete the form below and one of our specialists will contact you.
