By Sarah Grotta, Chief Fintech Officer, FinWise Bank
For most of the history of payments, we have designed systems around a fairly basic assumption: somewhere in the transaction, a person is making a decision.
A consumer swipes a card. A business approves an invoice. Someone authorizes an ACH transaction or initiates a real-time payment.
Artificial intelligence is beginning to challenge that assumption.
I recently had a fascinating conversation with Alex McDougall, CEO of FUTR, on our EYE ON Fintech podcast about the rise of AI agents in consumer financial services. One of the ideas we explored was the transition from AI that simply recommends an action to AI that can actually execute it. Alex’s view was that agentic commerce can be incredibly powerful when AI has sufficient context, while money should move within the permissions established by the consumer.
That raises an important question for the payments industry:
What happens when the entity initiating a payment is software?
From human-initiated to machine-initiated
This is quickly becoming more than a theoretical exercise.
Visa recently described two emerging forms of agentic commerce. In one, an AI agent acts on behalf of a person to do things such as book travel or manage subscriptions. Those transactions can look much like today’s e-commerce payments and may work comfortably on existing card rails. [visa.com]
The second model is considerably different: software paying other software. An AI agent might purchase an API call, data, computing capacity or another digital service. These transactions can be extremely small and extremely frequent. Visa notes that the economics become challenging when the average transaction is a fraction of a cent and a fixed card fee could exceed the value of the transaction itself. [visa.com]
That is where the payments conversation gets interesting.
Our existing rails are remarkably good at what they were designed to do. But machine-to-machine commerce may introduce different requirements around frequency, economics, identity, authorization and the definition of who, or what, is actually making the payment.
Authorization may become more important than speed
On the EYE ON podcast, Alex McDougall and I discussed an issue I believe will become central to this evolution: Where should an agent’s authority stop?
Today we often think of payment authorization in transactional terms. With an autonomous agent, authorization could increasingly become policy or rules driven.
A person or business might authorize an agent to act within defined parameters. The difficult questions then become: What exactly did the user authorize? Does the agent have permission for this particular transaction? How do the other parties know that? And when should an action come back to a human for confirmation?
These are issues the industry is actively working through. Mastercard, for example, identifies clear user intent, secure credentials and verifiable agent identity as important elements of agentic-commerce protocols. [mastercard.com]
The innovation, in other words, cannot stop at teaching a machine how to pay. We also need reliable ways of establishing who authorized it to pay, what it has permission to do, and under what conditions.
Do we need new rails, or smarter use of existing ones?
I don’t think the answer is necessarily replacing today’s payment infrastructure.
We are more likely to see different rails and protocols used for different types of agentic transactions.
Visa argues that agent-driven consumer transactions may continue to fit comfortably on card rails, while machine-native micropayments are creating different economic considerations. Its research with Artemis examined newer protocols including x402 and the Machine Payments Protocol, both being used for machine-native transactions. [visa.com]
At the same time, existing payment networks are adapting. Visa is developing capabilities around agentic commerce through Visa Intelligent Commerce and its Trusted Agent Protocol, while Mastercard is working on Agent Pay and interoperability with emerging AI-commerce protocols. [corporate.visa.com], [mastercard.com]
The question therefore may not be which rail wins.
It may be how financial institutions create an environment in which multiple payment methods can coexist while maintaining clear authorization, identity, controls and accountability.
The bank still matters
There is an irony in all of this.
As commerce becomes more automated, the need for trusted financial infrastructure may actually become greater.
AI can make decisions faster. Software can initiate transactions automatically. Payment systems can move funds almost instantly.
But someone still needs to establish the rules under which money moves.
At FinWise Bank, we’ve spent considerable time thinking about modern money movement. Our MoneyRails platform supports payment methods including ACH, wire, RTP, FedNow, Mastercard RPPS and Visa Direct. That experience reinforces a broader point: innovation in payments isn’t simply about adding another rail. It is about determining how money can move appropriately for a particular use case.
Machine-to-machine commerce may test that principle in entirely new ways.
I don’t believe we know exactly what the machine economy’s payment infrastructure will ultimately look like. But the questions financial institutions should be asking are becoming clearer.
Who is initiating the transaction? Who authorized it? What are the limits? Which rail makes sense? And what happens when something goes wrong?
As we have seen countless times in the past when we have new capabilities, regulations that fit the need will not surface until there is clarity around the risks. There is always a lag before the guiding principles are established. It will be up to those who are early trailblazers to shape the parameters and ensure the environment is secure for all participants.