Yapily CEO, Stefano Vaccino recently joined speakers from Google and NVIDIA to talk about where agentic commerce is heading, exploring how open banking can help answer questions about data, consent, and trust.
Since the start of online payments, the buyer at checkout has been a person: someone compares options, considers the price, and then decides whether to click “buy”. But soon, human buyers will be joined by AI agents acting on someone's behalf, buying with a mandate and a budget rather than a browsing habit.
One projection puts agent-driven commerce at between 200 and 400 billion within three years. And it’s this kind of potential that has the largest technology companies accelerating the building of tools that agents will use to start buying.
This article explores four key themes from the discussion, examining which parts of agentic commerce are close to ready, which are not, and why open banking might offer solutions to these problems.
1. Merchants need to be ready for a profoundly different buyer
When an agent does the buying, it arrives with its instructions already set: buy this type of product, within this budget, to this specification. Emma Zaraisky of Google described this as a move towards high-intent purchases, where the agent turns up with a clear mandate, and the merchant's job is simply to be ready for it.
For merchants, this is a genuinely different challenge. Instead of designing a page to persuade a human to buy, you have to optimise it so an automated buyer can find your product, assess it, and complete the purchase with no person involved. When the buyer is an agent, your payment rails and product data have to be built for machines to read. That prompts some early and awkward questions: if agents become the buyers, what are merchants actually designing their ecommerce pages for? And how do you create these pages for agents without diminishing the experience for humans?
"You have to be ready when that agent is there."
Emma Zaraisky, Google
2. An agent is ultimately limited by the depth and quality of data it can reach
Most of the excitement about agentic commerce focuses on intelligence, meaning models that can reason, plan, and act on their own. An agent without real-time, in-depth, and accurate financial data, though, can do very little beyond hold a conversation and advise without any certainty. Without that level of data, it cannot tell you what you can afford, which account to pay from, or whether a purchase makes sense given your complete financial picture, your spending habits, and your projected cashflow.
Jochen Papenbrock from NVIDIA explained how the intelligence layers are being built to solve this. Models trained on transaction, event, and customer data that agents can draw on to make better decisions are already being rolled out with partners including Revolut and Mastercard. Stefano drew out the consequence, explaining that an agent is only as “intelligent” as the data it holds about you, because its reasoning can only ever be as good as the information it has to work from.
This is the point where open banking offers agentic commerce something tangible. Open banking data (account information access), granted by the user through a regulated framework, is one of the few ways an agent can connect to accurate, up-to-date financial data at the scale agentic commerce will demand. Building the intelligence and supplying the data are two separate jobs, and neither works without the other. That leads to a practical question the industry has barely started to answer: who controls access to that data, and can today's model cope when millions of agents are requesting it at once?
"An agent is only as smart as the data it has about you." Stefano Vaccino, Yapily
3. The question of trust and consent is yet to be satisfactorily answered
Autonomy gets most of the attention when it comes to agentic commerce, and in theory, agents can already act on your behalf. But in practice, the question is whether you would allow it to, and under what conditions?
In many ways, open banking has already done much of the work that the rest of the industry is only starting to explore. Stefano highlighted that the consent model that open banking runs on is already very close to the building block agentic commerce needs: regulated, real-time, limited in scope, and fixed in duration. It means a user can give a provider permission to carry out a specific task on their behalf, for a defined period, and can withdraw that permission whenever they choose.
Jochen Papenbrock put the autonomy side into everyday terms, comparing an agent that acts for you to a car that drives itself. Letting an agent act is the part we can already do; letting it act with your full trust is the part still being built, and consent is at the heart of it.
"These consents are real-time, in scope, and defined in time."
Stefano Vaccino, Yapily
4. The fragmentation problem no single company can fix on its own
The clearest obstacle the discussion raised was fragmentation. There are many competing protocols, many overlapping standards, and a set of experiences that still do not join up cleanly.
Emma Zaraisky called this an inflection point, where the industry has yet to agree on which protocols to build on, and where the final stretch towards a genuinely usable experience has not been reached.
Stefano laid out the challenges he expects the industry to work through:
Agreeing how consent is expressed
Defining the mandate an agent operates under
Establishing who is liable when things go wrong
Reducing the fragmentation between systems that currently do not speak the same language.
He pointed out that if every platform speaks a different language, agentic commerce will not work. The encouraging sign is that the biggest players seem to recognise this and are beginning to line up behind shared standards rather than building competing ones. Whether that agreement holds, or whether the fragmentation simply reappears one level up, is still an open question.
"If all platforms speak a different language, it will never be a success."
Stefano Vaccino, Yapily
The compute bill behind agentic commerce
Running agents at scale is expensive, and that cost will determine which companies can afford to participate.
Nvidia’s Jochen Papenbrock described the infrastructure behind agents as an AI factory, where what matters is how efficiently raw compute is turned into useful decisions, and where the total cost of running that infrastructure has to be planned for from the start. Every guardrail an agent checks against, every decision it grounds in data, and every policy it validates consumes compute. As agentic commerce grows from a novelty into something people use every day, that running cost will become more of a strategic decision rather than a technical detail. It raises the question: will the cost of running agents at scale decide which companies get to compete?
"The system for the agents is an AI factory"
Jochen Papenbrock
Where this leaves us
The parts of agentic commerce that make the demos so impressive, like the reasoning, the autonomy, and the intelligence, are the parts closest to being finished. But the parts that will actually decide whether it works in real-world scenarios, like how an agent gets access to your data, how consent is granted and limited, who carries the liability when an agent makes a mistake, and whether the industry can settle on a shared language, are still yet to be built.
We do not know which protocols will emerge successful in the current race to dominate, how liability will be handled when an autonomous buyer gets something wrong, or whether a consent model built for humans can survive constant, agent-driven activity. Even the development timeline is a best guess; the panel suggested mass adoption is two to three years away, but predictions in this field have a habit of slipping.
What stood out most from this conversation was the need for a clear understanding of how the ecosystem that enables agentic commerce will work. Companies writing the protocols and the ones building the compute might be diving from the front, but there is a need to recognise that in the background, companies like Yapily needed to provide the data connectivity, payment initiation, and the consent framework layers for agentic commerce to operate in the real world.
Open banking already securely connects consumers to thousands of banks across Europe, and it already offers a working, regulated model for consent. Pay by Bank already moves money directly between accounts, and Bank on File can provide the guardrails necessary for agents to act within strict controls. That does not resolve every question agentic commerce needs to answer, but it does mean the groundwork for some of the most difficult parts is far more advanced than that many in the industry might realise.