Imagine you are looking for a good bottle of wine as a thank-you gift for the guest speaker at your event. Instead of searching yourself, you give an AI a task: find a nice Italian red wine for around EUR 50, check reviews, and order it from the cheapest available supplier. The AI compares options, monitors prices, places the order, and completes the payment. You only define the conditions.
The above example captures the essence of agentic commerce: AI that not only advises or acts as a sparring partner, but executes the entire commercial process, including payment. According to forecasts, approximately 45% of e-commerce transactions are expected to be handled by agentic AI by 2027.
The FinTech team at Kennedy Van der Laan discussed these developments – and the associated risks – with Hugo Reijkens, Director Digital Payments at Mastercard. Mastercard is among the parties actively working on the infrastructure required to enable these types of applications at scale.
From Single Purchases to Autonomous Action
“Much of what is currently being built consists of relatively well-defined use cases,” he explains. “Think of an agent that selects a product and makes a single purchase from a merchant. Those are the first steps that we are now actually developing and testing.”
At the same time, according to him, this is only the beginning. “If you take this further, the possibilities are effectively unlimited. Agents that continuously monitor, combine multiple providers, and execute transactions based on predefined conditions. For example, an agent that tracks airline ticket prices and automatically books once a certain threshold is reached, or that independently makes purchases when specific criteria are met.”
“The challenge today is not so much what is technically possible, but how we can scale this in a reliable way,” says Reijkens.
Why Friction Still Exists Today
According to Reijkens, there are roughly two reasons why agentic commerce has not yet truly taken off.
“First of all, there is simply no visibility into who is acting,” he says. “Within current payment flows, you cannot see whether a transaction is being carried out by the consumer themselves or by an agent acting on behalf of that consumer.” At first glance, this may seem like a technical detail, but it goes to the heart of the system. The current payment infrastructure is built on the assumption that a human initiates the payment and is directly involved in it. Once an agent takes over that role without this being explicitly visible, an important part of the context disappears. Transactions become harder to understand, more difficult to monitor, and ultimately more challenging to assess from a risk and liability perspective.
“On top of that, today we mainly see the outcome of a payment, not the step that precedes it,” Reijkens continues. “We can see that a payment has been made, but not what exactly the user instructed the agent to do.” That step becomes essential in agentic commerce. After all, the user does not provide a traditional checkout instruction, but rather a mandate containing conditions, ranges, and choices that the agent must make. If that underlying intent is not explicitly recorded, disputes can quickly arise when the outcome differs from what the user had in mind.
“These two points are closely connected,” says Reijkens. “As long as it is unclear who is acting and on the basis of which instruction, it remains difficult to make transactions truly explainable and controllable.”
Who Pays the Bill?
From a legal perspective, the key questions lie precisely at the intersection of intent, authorisation, and execution. Who bears the risk if an agent acts within the given parameters, yet still produces an undesirable outcome?
Existing safeguards such as Strong Customer Authentication (SCA) are also expected to continue playing an important role within agentic commerce. Not only in authenticating the payment transaction, but also in recording the user’s intent. At the same time, this raises new legal questions. The current PSD2 framework assumes a relatively direct relationship between the user and the payment. When an agent acts on behalf of the user, it becomes increasingly relevant to determine how that original instruction relates to the final payment and at what stage authentication should occur.
As a result, the core question becomes broader: how do we allocate responsibilities among the consumer, the agent provider, the merchant, and payment service providers in a model in which not only humans, but also systems, can act? It is precisely at this point that it will become clear whether the existing framework is sufficient, or whether new rules of the game are required.
