Every year, our industry finds a new shiny concept to get excited about. It dominates conference panels, appears in every other LinkedIn post, and quickly becomes something you’re expected to have an opinion on. Right now, that shiny concept is ‘agentic payments’.

We’ll be honest, our initial reaction was cautious. Not because the concept lacks merit, but because there is a gap between the current understanding of how it works and the practical reality. 

At its simplest, the idea is straightforward. You instruct an AI agent to complete a task on your behalf that requires a payment. For example, you might want to find a product, compare options, and make the purchase. Typically, the ‘agentic’ part is the research, decision-making, and execution culminating in the purchase. The ‘payment’ part is simply the final step in that process. It is very similar whether a human is making the payment or if it is being initiated by an AI agent. Which raises the obvious question. What is actually new here?

While we would not recommend it,  when someone gives their card details to another person and asks them to make a purchase on their behalf, the outcome is broadly the same as entrusting an AI agent. With agentic payments, the mechanism hasn’t fundamentally changed, only the actor has. We’ve simply swapped a human proxy with a digital one. So why all the excitement?

Proponents of agentic payments will tell you the innovation lies in how the transaction is authorised and executed. Instead of manually entering card details or triggering authentication, the idea is that the agent operates with a form of delegated authority. The agent presents a (hopefully tokenised) credential, the merchant recognises it, and the transaction proceeds.

In theory, this creates a smoother, more automated experience. In practice, it raises a set of questions that are far more interesting than the technology itself. The first is around control.

If an agent is acting on your behalf, how do you ensure it does exactly what you intended? As consumers, we would still want a moment of confirmation before a payment is made. Not because we don’t trust the technology, but because we understand how easily small errors can become expensive ones. After all, automation without control may solve one problem but in turn creates a new, significant risk.

The second question is about responsibility. This is where the industry conversation becomes significant. If you delegate your payment authority — whether to a person or a machine — you are still accountable for what happens next. 

What changes with agentic payments is not the principle, but the scale and the speed at which things can go wrong. There’s also the risk that in many cases, you could be stepping outside the protections that payment schemes, banks and even potentially regulation typically provide or have anticipated. The moment you start sharing credentials or delegating authority, you begin to erode those safeguards and probably break terms and conditions of the payment instrument. It might be a minor technical issue, but it’s certainly a greater legal and commercial one.

The third, and in our view, most important question is, what problem are we actually solving?

When you strip away the terminology, most of what is being described is simply a new use of existing capabilities. Discovery, decisioning, and execution brought together in a single flow. That’s useful, certainly, but it’s not a reinvention of payments. 

What about using existing failsafes to limit new risks? 

For example, instead of giving an agent unrestricted access to a funding source, you issue a dedicated, programmable payment instrument. A card or wallet with predefined limits, conditions, and permissions. An environment in which an agent can only operate within specific constraints designed by the user. This is much like the controls you create around card use within any solid fintech product.

Moving from blind delegation to controlled automation is not hypothetical; the building blocks already exist. Tokenisation, virtual cards, spend controls, and merchant restrictions all exist in the current payments infrastructure. The difference is how they are applied to enhance the agentic solution.

You could even extend this thinking further. In a recent internal discussion, we touched on the idea of programmable stablecoins as a potential model for agentic commerce. If money itself carries rules in terms of where it can be spent, under what conditions, and for what purpose, then the risk no longer sits solely with the agent. The constraints are embedded in the asset as a second layer of defense.

It’s a similar story to what we’re seeing more broadly in payments innovation. Stablecoins, for example, promise speed, programmability, and direct ownership of value, but their adoption depends less on capability and more on trust, regulation, and behaviour change.

Agentic payments sit in a similar place. There is something here, but it’s likely not what the headlines suggest. It’s not a new payment method. It’s not a fundamental shift in how money moves. It’s an evolution in execution, which necessitates a procedural, cultural and conceivably regulatory shift to enable a new, potentially very useful solution to thrive. 

Like any evolution in our space, mainstream adoption will depend on something far less exciting than the technology itself. Consumers don’t adopt ideas, they adopt outcomes they understand and trust.

Until agentic commerce proves it can deliver that with clarity and control, it isn’t the future of payments; it’s an interesting layer on top of what already exists.