An AI agent was ready to pay. It called Nuvei

Dwayne Gefferie [00:00:00]:
Good afternoon, everyone, and welcome. I am Dwayne Gefferie, payment strategist and columnist at The Paypers, and I will be moderating today. Before I introduce the speakers, I want to have a few quick notes on timing and logistics. The discussion will run for approximately 45 minutes and we'll keep about 10 to 15 minutes for Q&A. On the right side of your screen, you can see the chat panel, which includes a chat room and a Q&A widget. Keep your questions coming throughout the session — we will get to as many as we can. At the end of the session, subtitles will be available; you can activate them by clicking the CC icon at the top right of the webinar room, just to the left of the chat tab. Subtitles are available in English, German, French, Spanish, Italian, and Portuguese. The recording will be available afterwards, so don't worry about taking notes. So let's get started. Today's topic: agentic commerce is full of vision, roadmaps, and demos. But what this panel actually did is process a live agentic transaction in production with a real merchant and real issuers. So rather than another conversation about where all of this is going, we are going to open up the hood on how it was made to work, what broke along the way, and what merchants should take from it. So let me hand it over to the panel to introduce themselves. Mentxu, would you like to start us off, in about a minute — who are you, and what was your role on this particular pilot?

Mentxu Triviño [00:01:39]:
Thank you very much, Dwayne. I'm Mentxu Triviño, I'm VP of Payment Partnerships. My role sits at the intersection of networks, merchants, issuers, banks, technology providers, and payments infrastructure, which gives me a front-row seat to how commerce is evolving. Over the past year, I've had the opportunity to work very closely with Adrian Trache from Visa and Andreas Moos, exploring one question that's becoming increasingly important to me: what happens when an AI agent — a machine acting on behalf of a human — is ready to make a purchase? That question ultimately led us to one of the industry's first real agentic payment proofs of concept, together with Visa, with Kings & Priests as the merchant, and with different European issuers. That's the topic I'm really excited about. While everybody's talking about how agents will shop, I'm focusing on what happens when they're really ready to pay.

Dwayne Gefferie [00:02:52]:
Thank you. Adrian, could you please introduce yourself — who are you, and what was your role in this pilot?

Adrian Trache [00:03:01]:
Hello. I hope you hear me well. My name is Adrian Trache, I'm Head of Product for Southern Europe. I'm leading the product team for Southern Europe, focusing on all product-related topics for Visa — and, of course, one of the latest products I'm very excited about is everything we're doing in the agentic commerce space, everything related to agentic commerce and payments on the Visa rails. I'm super excited to have been part of this pilot, working very closely with Mentxu. What we're focusing on at Visa, as you well know, is that we want to be the best way to pay and be paid, and that's what we're looking to achieve as part of this pilot and this project.

Dwayne Gefferie [00:03:49]:
Perfect, thank you very much. And Andreas — last but not least, please introduce yourself and your role in this pilot.

Andreas Moos [00:03:56]:
Thank you very much, Dwayne. My name is Andreas Moos. I'm a strategy and business development manager focused on consumer goods. Our role was to orchestrate everything besides the payment itself — so when you look at agentic commerce processes, everything that happens before and after the payment was our role. That's product information, customer information coming together so the customer has a great experience with an agent — all these commerce processes were ours to orchestrate.

Dwayne Gefferie [00:04:37]:
Fantastic, thank you very much for that. So just to give everybody an idea of where we're going: first we'll talk about where the market actually is today, then the pilot itself — what happened, what it proved, and how hard it was to execute — and then what has to be true for all of this to actually scale. Along the way we have two polls, so we'd love your input to get a perspective from yourselves as well. So to start, Mentxu — where are we now on agentic commerce, and where are we headed?

Mentxu Triviño [00:05:18]:
That's very interesting, because as payment experts with long experience in the e-commerce ecosystem, we know the payments industry has spent decades learning how to trust people. For me, the next challenge is learning how to trust machines — because we're entering a world where a machine may act on behalf of the customer, not only helping that customer discover products but ultimately making purchases on their behalf. That changes the conversation completely, and it's been part of the conversations I've had with both Adrian and Andreas. At this point, we're no longer just talking about AI — and that's important to clarify from the very beginning: we're talking about payments, and why we're talking about payments is because we're also talking about trust, and about the ecosystem that has to support both how payments are committed and how we build trust along all the stakeholders. That's why I believe agentic commerce becomes truly strategic at this moment, where the agent is ready to pay — and we proved that it's able to pay. So the big challenge of today's conversation is thinking about how we learn to trust machines. If we think about the evolution, it's actually very simple: first, AI helped people find things — what I call the discovery phase. Second, agents help people buy things. Third — which is the most important one, and what everybody's expecting — agents will buy things on behalf of people. That's the progression we're seeing today, and that's the infrastructure we're building today, which is why payments infrastructure becomes really central. Because once an agent can pay on behalf of the customer, the ecosystem needs to answer questions we tackled during this proof of concept: which agent is acting, who is the agent acting for, under what authority, and can the payment be trusted? That's why I often summarize it this way: AI helps people decide, agents help people buy — but as soon as they start buying for people, payments must learn how to trust these machines.

Dwayne Gefferie [00:07:49]:
You mentioned trust, and I think trust is probably one of the key factors in all of this. Your own consumer research suggests trust isn't there yet. What are consumers actually saying they would need before handing over that buy button?

Mentxu Triviño [00:08:08]:
That's precisely what we built with this proof of concept, which goes beyond a demo. From a technical point of view, the technology already provides this agentic payments infrastructure. When we think about trust from the customer's point of view: yes, they're using agents for discovery, yes, they're using agents to get better information, comparing pricing, everything related to the pre-purchase experience — and they're eager to use it. But the question of trust becomes very relevant at the moment of payment. I might go to an open agent, a merchant agent, or a third-party vertical agent, and the moment that agent asks for my payment details, my payment credentials — that's where trust matters. Adrian can build on this based on other Visa research, but typically, consumers trust the brands they interact with. Agents are all new. So we still see that customers will need more trusted procedures, and that's where payment infrastructure plays a great role — because we can ensure customer identity, customer mandate, clear agent identity, and all the fraud and risk management procedures that already exist, which will trigger transactions toward the agent. For me, it's just a matter of building that final layer.

Dwayne Gefferie [00:10:06]:
Absolutely. And of course there are two sides to it — consumers doing the shopping, and merchants on the other side. There seems to still be a gap between what merchants say they're doing and what they're actually doing. What are you seeing in the market, beyond this pilot?

Mentxu Triviño [00:10:29]:
I have almost daily conversations with merchants, e-commerce platforms, and other stakeholders. From my point of view, merchants clearly know what agentic commerce is, and many are already doing things in this space. There's a first layer of merchants asking us what to do and when to start. There are also merchants starting to incorporate identity frameworks or asking questions — how will agentic payments impact my customer journey, how will I identify the data, how will this impact my customer relationship? Those questions are already being raised by early adopters. We have merchants working with us running pilots like this one. We're also interacting with merchants a step further along, considering piloting. And of course, we have big platform merchants — the Amazons of the world — already integrating this into their operating systems because of the scale they have. So I'd say there are roughly four stages of merchants, and our main role is to educate the industry, help them ask and answer the questions they have, and go hand in hand with the different stakeholders — because this is an ecosystem play.

Dwayne Gefferie [00:12:02]:
Thank you. I think that gives us a good basis to continue. I also want to get everyone involved — we have our first poll: where does your organization sit today in agentic commerce? Not engaged, exploring, piloting, or operating. While everyone answers — Andreas, how are you already working with agentic commerce today?

Adrian Trache [00:12:51]:
So, how are we approaching agentic commerce today? I'll start with some history. We introduced agentic commerce from Visa's perspective about a year ago, when we made the first announcements of Visa Intelligent Commerce. A lot has happened since. One key topic when we introduced the framework is looking at how innovation has shifted over the past two decades — from e-commerce to mobile commerce to social commerce to in-app commerce, and now agentic commerce. One thing is certain: customers want trust, security, and scale. The payment needs to work frictionlessly, invisibly, behind the scenes — reliable and always on. So what we've done as Visa is put in place a framework — a set of rules and regulations for how agentic commerce can operate. It's important to see that there's a new player at the table: the agent. At a high level, everything has shifted from the consumer initiating the payment to the consumer delegating to an agent — "do this payment on my behalf." That brings a new player to the table, which is why we put in place a framework, a set of APIs and functionalities, and Mentxu will elaborate later on how agents can operate on the Visa network securely and in a trusted way. That's the key point of agentic commerce — though trust, of course, is much broader than just onboarding the agent.

Dwayne Gefferie [00:15:37]:
Fantastic, thank you. We're getting good feedback on the poll — almost 67% say they're in the exploration phase. Andreas, for people in that exploring stage, what actually moves them toward piloting?

Andreas Moos [00:16:07]:
I think what's really important is data. There was a time when it was crucial for a merchant to do the homework — checking product data quality, API readiness. Now it's not just important, it's crucial; without it, you can't access this agentic space at all. The agent needs to be able to give information on prices, stock, orders, logistics, and help with discovery — so it needs to know your product. Product information and data quality are critical. Now everything comes together: how do you give the agent access to your different data stores so it can give the best experience? Trust, as Adrian mentioned, is important here too. From my perspective, autonomy for an agent isn't a switch — yes or no. It's more like a dial. Depending on your product and where you are right now, you don't have to go straight to a fully autonomous agent; you can start at a smoother level. But data is the most important thing — data errors become more costly, so quality has to be very good.

Dwayne Gefferie [00:18:15]:
Perfect, thank you. It's important that people understand: going from exploring to actually piloting means having your data in order, having that structure ready to provide agents with relevant data. Andreas, staying with you — the agent has become a new point of purchase. What specifically changes for a merchant when that point of purchase moves from their own checkout into an agent?

Andreas Moos [00:18:47]:
If it's an agent on your own site, you have to check product data quality and API readiness — is everything there so the agent can give an answer? Do you have direct customer access, zero- and first-party data that lets the agent learn and give the best answer? Differentiation and trust — reviews, certifications, measurable brand references, strategic signals — are now very important. And if the agent is the new point of purchase — not just on your own site, but for example through Google's UCP, this becomes even bigger for retailers. There's a loss of direct customer visibility: the customer no longer necessarily sees the shop, the listing, or the ad the way they used to. Now it's the first answer that's findable that matters — so you have to be visible; this "GEO" (generative engine optimization) topic is super important — how do you go from findable to selectable? For commodity products especially, without differentiation or trust signals, the agent decides purely on price, availability, or delivery time — which you don't want. So trust and differentiation signals matter, so the consumer knows exactly what your brand stands for. There's also platform dependency — always has been — but now the agent platforms have new rules, new gatekeepers: ChatGPT, Google AI Overviews, Perplexity. You have to learn their rules — how to be findable and selectable there. And, again, data errors become more costly: incorrect stock, prices, or attributes can lead to bad decisions and bad returns. In a nutshell: zero-defect data is the priority, whether it's an agent on your own site or on a third party like Google.

Dwayne Gefferie [00:21:49]:
Absolutely, thank you. Adrian — card networks have absorbed new channels before. In the last forty to fifty years, we've gone from in person shopping with POS devices to ecommerce, mobile wallets. What makes agent traffic different from anything the network has handled before?

Adrian Trache [00:22:18]:
Coming back to the previous point: what's essentially different and you mentioned very well - and I'm speaking about the online space, ecommerce, mobile commerce, social commerce, and so on, mobile wallets. In agentic commerce, the AI agent does the search, the discovery, evaluates, and initiates the transaction on behalf of the cardholder. All the previous channels were essentially just additional payment channels — but here, agentic commerce actually changes who participates in the transaction, and the new participant is the agent. This changes the traffic and creates new challenges. Until now, the question was: as a consumer, can I trust this card? As a merchant, can I trust this consumer or this card? We see a lot of fraud trends around this. But in agentic commerce: can I trust this agent? Can the agent trust the merchant? And from the merchant's perspective: can I trust the agent — is it actually operating legitimately on behalf of the consumer's intent? So we need to adapt and make sure that, as a network, we understand who the agent is, who it represents, and what it's actually allowed to do. That's the key difference.

Dwayne Gefferie [00:24:33]:
Fantastic, thank you. I want to move to the next step — we've explored where we currently stand, and some of the challenges around trust, control, and identity. Now I want to dive into the pilot itself. Mentxu, could you walk us through the live transaction — what happened, and, most importantly, why was it so hard to execute?

Mentxu Triviño [00:25:02]:
Yes — if you allow me, I'll make this personal for a moment. When Adrian and I first discussed this, almost a year ago (or more) at an industry event in Ibiza, the opportunity was clear: could we show that agentic payments weren't just a concept, but something that could work across real payments infrastructure? Then Andreas and a colleague joined, which was critical, because you need the commerce infrastructure perspective in the room. One key learning for me — and I believe for all of us — is that you cannot design agentic payments only from a payments perspective. You need the merchant, the network (Visa), the acquirer (a role Nuvei can play), the issuers, and the technology partner all lined up. That's what made this proof of concept matter to me — it wasn't one company showing a demo saying "look, the technology works," because technology always works. I've been in payments since the early days of the internet. It's always about the ecosystem, always about putting the customer at the center — in this case, the consumer. Nuvei went one step further by developing our own agent to speed up time to market for this proof of concept. My personal experience wasn't about difficulty — it was that we were learning by doing, and everything was completely new. There was nothing to look at for reference; we just had to come up with solutions. The energy and talent that all these companies brought was amazing — it kept us moving and made clear we had something special in our hands. No one can do this alone. With the right talent behind an infrastructure challenge, you can make it happen.

Dwayne Gefferie [00:27:34]:
And I believe it had to clear across multiple European issuers — why did aligning them matter so much?

Mentxu Triviño [00:27:49]:
I'll pass this to Adrian, since that's part of the contribution he made to the project.

Adrian Trache [00:27:56]:
Sure, thank you. To build on what Mentxu said — in any innovation effort, there are generally three phases: exploration (the theoretical part, the rules, the flow, how it will work), doing the thing (the pilot — exactly what we did), and then scaling it up. Exploration is fine, we did that, and then we did the pilot. But for the pilot to be relevant, we didn't just want to make one transaction and see "can we do an agentic payment transaction" — we already knew we could do that. We wanted to test, in a controlled but still live environment, issuers from different countries in Europe, as well as outside Europe, and see how they interact — from a regulatory perspective, from an ecosystem perspective, and how the transaction would actually succeed. That's why it was important to say: this is a pilot — not generally available to all cards in the world, but available to a relevant mass of cards and consumers who could test it, with real-life issuers and real-life merchants, everything processed through Nuvei and through Visa at the network level.

Dwayne Gefferie [00:29:58]:
Absolutely. So what did this prove that a slide deck or a sandbox couldn't prove? What did you really learn from this entire process?

Adrian Trache [00:30:17]:
I'd highlight three things. First, making sure the ecosystem actually works — as Mentxu mentioned, this is an ecosystem effort. We wanted real consumers, real cards, real banks, a real agent — no mockups, no demos, no HTML simulation — a real agent doing the work with a real merchant actually completing the purchase. That's checked. Second — as with any real-world example, you discover new friction points. This let us validate that the agent is securely tokenizing the credential, securely doing authentication and authorization, and all the other operations — and that it's actually working. That comes back to trust and security: it needs to be in place, and it needs to work frictionlessly, not "sometimes it works, sometimes it doesn't." And last, but not least: this proves it's not just a future concept — it's an actual, functioning thing. We can now move from "can this work" to "how do we scale it."

Dwayne Gefferie [00:32:01]:
Andreas, from the other perspective — working closely with the merchant on the technology implementation — what did the shopper actually experience, and what was happening behind the scenes that they never saw?

Andreas Moos [00:32:18]:
It's easy to say the shopper experiences a completely new journey, with an agent by their side guiding them from discovery through execution. That's the bridge — and with this purchase capability, a whole new set of possibilities opens up. A couple of examples: if a product is out of stock, the agent can now let the shopper know when it's back in stock — through new channels, like WhatsApp, not just its own platform. So instead of just "your product is back in stock," it becomes "your product is back in stock — can I just buy it for you?" That turns an out-of-stock moment into a pre-qualified demand pipeline, which is new for both the customer and the merchant. Another example: a "price agent" that notifies you of a discount or promotion, or when a product becomes available in your size, and can act on the rules you've set.

Dwayne Gefferie [00:35:35]:
Absolutely — I share your enthusiasm. It opens up so many new possibilities from both a shopper's and a merchant's perspective, in terms of new products, new services, and new pipelines you can activate through agentic commerce.

Andreas Moos [00:36:07]:
One last thing on trust signals: starting with smaller use cases — a "back-in-stock agent" or "refill agent" for something like pet food you need every four weeks — you can tell the agent "just do it with my rules." I think this builds trust step by step. It's less about full autonomy right away and more like a delegation agent or concierge working for you. As merchants run these use cases together with us, trust builds month by month, because they see it actually working — and from there, autonomy increases.

Dwayne Gefferie [00:37:07]:
Absolutely, I fully agree. Now, our second poll: what's your top concern about agentic commerce? Options include architecture and protocol choice, trust and fraud, customer relationship and data, and commercial readiness. Mentxu — from your conversations with merchants, what concerns are you hearing?

Mentxu Triviño [00:37:52]:
I think there's a rational discussion happening here. On architecture and protocol choice — no one wants to bet on the wrong standard too early, given how many protocols are emerging; no one wants to commit to one and then switch in six months. That's where enablers like Nuvei play a role, building a protocol-compatibility layer so merchants stay compatible regardless of which protocol an agent uses. On trust and fraud — a fair, recurring question, since both of our current models are built around trusting humans, not machines. That's exactly the conversation we need to keep having, building behavioral mandates, permissions, and risk frameworks, as Adrian mentioned. On data — I think Andreas made a strong point about this being one of the most strategic concerns: the customer relationship around data. Part of our role is helping merchants keep control of that. For us, we're on the side of the merchant even as we're on the side of the agent — but the customer relationship needs to stay with the merchant.

Dwayne Gefferie [00:39:44]:
Looking at the poll results, we actually have close to a three-way tie. It surprises me that customer relationship and data isn't a bigger concern — Andreas, that's probably good to see, that people see it more as something to handle than a blocker. But trust and fraud remains a major concern, along with commercial readiness. Adrian, from your position in the middle of this — working with Nuvei, acquirers, issuers — what are the top concerns you're seeing?

Adrian Trache [00:40:39]:
I'd say largely the same as what the poll shows. I'd actually link commercial readiness to architecture, protocols, and trust and fraud — if there's trust, no fraud, and there's scale and interoperability, then yes, it's commercially ready from the merchant's perspective. On the issuer side, we're addressing their concerns using the same tokenization infrastructure, enhanced to identify the agent — so credentials transact securely and safely. From Visa's perspective: yes, there are multiple protocols out there, some competing, some operating at different stages of the transaction.

Dwayne Gefferie [00:42:47]:
Thank you very much — I love the feedback we're getting on the polls. Before we move to our last segment, if you have questions for any of the panelists, please put them in the Q&A — be as direct as possible. Now, on scaling: what needs to happen before agentic commerce can scale, and before truly autonomous payments are possible? Adrian, let's start with you, then Andreas, then Mentxu.

Adrian Trache [00:43:53]:
Building on the earlier point — we're seeing three forces shaping scale. First, innovation itself creates new possibilities: a new consumer experience layer driving an explosion in this space — better agents, faster discovery, smoother interfaces, more guided, easier to use, across multiple surfaces. Second, and critical: interoperability — how well the ecosystem lets different participants (issuers, cardholders, consumers, merchants) interoperate across multiple networks. And last — trust, again — the thing that lets both innovation and interoperability scale with confidence. Right now, the industry is moving ahead on innovation and on interoperability.

Dwayne Gefferie [00:46:38]:
Andreas, working with merchants and providing the technology — what do you think needs to happen before agentic commerce can really scale?

Andreas Moos [00:46:47]:
Adrian covered a lot of good ground, so I won't repeat it — trust is clearly central. When working with merchants, the first step is being findable in search engines — Google AI, and so on — through structured content and brand appearance. Then being selectable: product data, intent mapping, reviews, trust signals. Then executable: price, delivery time, order accuracy — everything needs to be true regarding that information. And last: is it governed — making sure agent autonomy is controlled, with guardrails. That's a big topic, and I think what Nuvei and Visa are doing here is great. Beyond that, I think what will change things a lot is the first wave of real use cases.

Dwayne Gefferie [00:49:13]:
Thank you very much. Mentxu, what do you think needs to happen before agentic commerce can scale?

Mentxu Triviño [00:49:21]:
For agentic payments to scale, trust needs to become infrastructure. We've spoken about different protocols — so when we think about infrastructure, what we're building is precisely this agentic payments infrastructure, providing a protocol-compatibility layer on the agent side — whether it speaks UCP, AP2, Visa Intelligent Commerce, or whatever protocol. What we need to ensure is that the payment infrastructure is able to translate across all of them.

Dwayne Gefferie [00:52:15]:
Well, thank you very much. We have a couple of questions from the Q&A. Adrian, this first one is for you: what type of new data must be inserted into an authorization or clearing request, per card scheme format, when the payment is initiated by an AI agent?

Adrian Trache [00:52:51]:
Thank you for the question. Yes, we're introducing additional data fields — as I mentioned, to identify the agent and other information, such as intent and mandate confirmation ("I am the one who authorized this payment"). I won't share the specific fields publicly right now, but I'm happy to go deeper with any customers directly. To summarize: yes, this leads to better and higher authorization and approval rates, and a better experience for both consumer and merchant.

Dwayne Gefferie [00:54:20]:
I fully agree. Everyone on the issuing or acquiring side, keep an eye out for those updates. Second question — probably for Andreas: should retailers be able to define their own trust policies for agentic commerce? Do you envision a risk-based delegation model where the level of autonomy granted to an AI agent varies based on transaction value, product category, or customer behavior?

Andreas Moos [00:54:54]:
On the first part — if it's your own agent, running on your own side, of course you can set your own policies. But we've already discussed why this matters: if your policies aren't the standard others use, and something goes wrong, trust is gone. That's a very important point.

Dwayne Gefferie [00:56:12]:
So essentially: would an AI agent be able to do more or less depending on transaction value, product category, or customer behavior — a repeat customer versus a first-time customer, for example?

Andreas Moos [00:57:39]:
I think this shows it really depends on the product and the use case. Searching for pet food is a lot less risky for me as a customer than, say, a higher-priced transaction — so it really depends on what your brand stands for, the product you're selling, and how you calibrate around that.

COLLAPSE ↓