Agents are coming to the checkout. Will anyone trust them when they get there?

At the Payments Innovation Forum's AI & Agentic Payments Forum on the 24 September, the industry swapped speculation for specifics. The technology has arrived, and fast, but the harder question is one of trust.

If you wanted a sense of how quickly agentic commerce is moving, Ed Leftwich of Dataiera supplied it in the last panel of the day. Visa and Mastercard had spent the afternoon describing much of this as months away. He was going live the following week.

That gap between roadmap and reality ran through the event. Issuers, networks, processors and fintechs spent the afternoon discussing one singular question: once AI agents do actually start buying on our behalf, how does everyone in the chain know they can be trusted?

The conversation was more practical than the hype might suggest. Nobody was debating whether or not this will happen. The conversation ranged across passkeys, token identifiers, spend limits, liability, and what happens when an agent buys the wrong T-shirt.

"Widely discussed… and yet, probably little understood”

PIF Executive Director Diane Brocklebank opened the day before handing over to Chair Jonathan Vaux, Head of Propositions & Partnerships at Thredd.

Vaux admitted he started in payments "in the late 1900s". In that time he has watched cards, e-commerce, mobile, tokenisation and contactless all change the industry. None of them, he said, created this much noise: "I don't really think there has been any innovation topic that has been so widely discussed… And yet, probably little understood.”

By the closing panel, the former Visa innovation lead would be making the slightly different point that a lot of this felt very familiar.

ANNA Money's heart transplant

ANNA Money has been a chat-first business since day one. Eight years ago, the team said, that chat ran on AI was "very stupid, kind of scary, but it would still work". Then the goal posts shifted. The panel compared it to Uber: a ten-minute wait for a cab once felt brilliant, and now three minutes feels slow.

So ANNA went for what it called a "heart transplant". The old system came out, and in went an AI that now handles onboarding, compliance questionnaires, Companies House filings, payroll and even retention calls. Some customers don't realise they aren't talking to a person. ANNA says 95% of conversations are now handled entirely by AI, with customer ratings a touch above what its human agents used to get. A single regulated accountant oversees the filing of almost 20,000 tax returns.

One might assume it was the AI model that took half a decade to get up to speed, but it wasn’t. The predictable deterministic models that were being paired with LLMs that can "make stuff up, just like humans do” was a clear problem initially.

"The human in the loop who broke things"

Then came the best story of the day. 

A customer's custom bank statement reached their auditors with the wrong dates and amounts. ANNA's first instinct was to switch the model off. When the team dug in, they found the AI had been right all along. A senior colleague had "corrected" it. It was, in ANNA's words, "the human in the loop who broke things". Some of its engineers would now happily take humans out of the loop altogether. The humans, unsurprisingly, would like more reviews.

The Martians have landed

PIF's newest member, DB Seret Group, arrived with the best nickname in the room. At the European Space Agency they're known as "the Martians". The name dates from when the agency found a team at the Vatican solving data problems with technology that had its roots in astronomy.

In conversation with Daniel Saliba, CEO of Lighthouse Compliance UK and a PIF Director, founder Emilia Di Bernardo, Riccardo Pianese and Petra Holusova made the case that agentic AI has a data problem hiding underneath it. Agents can only be as reliable as the records they draw on, and those records have a habit of being lost, corrupted or impossible to find. DB Seret has spent more than 20 years fixing that for the European Space Agency, the Vatican Library, NATO and Banca d’Italia.

AI will need quality data to give quality results

Di Bernardo described the European Space Agency project that began a decade ago: ten kilometres of paper and 230 separate document management systems spread across Europe. The answer was built on FAIR data (findable, accessible, interoperable, reusable), and she argued financial services needs the same discipline before it hands decisions to agents, noting that "AI will need quality data to give quality results."

Thredd: not all agents are going to be good agents

After lunch, Nuruz Zaman, Head of Risk & Digital Services at Thredd, gave the view from the issuer's side of the table. Before an agent spends a penny on someone's card, what does the issuer need in place?

He started with trust, which he said will be “foundational”, with the specifics following. Tokenisation and push provisioning, so agents get the right credentials with the right scope. Passkeys, which he called "very, very key": they authenticate the cardholder earlier, take friction out of 3DS, and carry the customer's intent through to any later chargeback. And a set of dials for issuers: agent allow and block lists, velocity and spend limits, MCC restrictions, and an off switch for individual agents.

That off switch exists for a particular reason. That reason being that "not all agents are going to be good agents," Zaman explained. Card-not-present fraud followed e-commerce, and fraudulent agents will follow agentic commerce. Thredd's answer is agent analytics, this means scoring how agents behave, much as fraud models score cardholders today.

Zaman was refreshingly honest about how differently issuers feel about all this: "You're going to have some issuers that simply don't want to be involved in agentic, which is understandable to an extent. You have some issuers that want to go full throttle.”

Thredd has spent the last six to nine months building out its agentic platform with networks and partners. It's also building an MCP server so back-end agents can help teams with fraud and chargebacks. Commerce gets the headlines, Zaman said, but the agents working quietly behind the scenes matter just as much.

Mastercard: the lights are on across Europe

Roza-Marina Mellou, VP, MACS & Agent Pay Product Management, Digital Payments Europe at Mastercard, promised to "demystify" the subject, and guessed the room had already heard the words "agentic commerce" more than 100 times that day.

By 2030, 30% of digital commerce transaction value is expected to be influenced by agentic AI

Her framing was a simple one. Commerce 1.0 happened in person. Commerce 2.0 arrived with the internet in the 1990s. Commerce 3.0 is agentic, and it's the first time in history that a human isn't the one directly making the payment. The first shift took almost 30 years, but this one is moving much faster. The numbers she shared backed that up. 58% of consumers have already swapped traditional search engines for Gen AI tools when looking for product recommendations, and 71% of those want AI built into how they shop. In Europe, 46% of consumers use AI tools day to day. And by 2030, 30% of digital commerce transaction value is expected to be influenced by agentic AI.

The first agentic transaction, a year ago, was "like opening Pandora's box", she said, because nobody in the chain could see an agent was involved. Was it legitimate? Did it buy the white T-shirt in size small, as asked, or ten of them in medium? Merchants had a problem too. They'd spent 20 years optimising their sites for search engines, and agents couldn't read their catalogues.

Mastercard's answer is Agent Pay, launched about 18 months ago. It rests on three pillars: Know Your Agent, which registers and certifies agents and adds agent identifiers to network tokens to "turn the lights on" for everyone in the chain; verifiable intent, an open standard launched this year that seals the customer's instructions together with the outcome; and interface standards that help merchants plug into agents. 

Agent Suite, which makes merchants discoverable by AI agents, and Agent Connect, a single layer across the different AI protocols, sit alongside it. None of it is reinventing the wheel, she stressed. 

It extends tokenisation and passkeys the industry already runs. And it's live: since February 2026, every issuer in Europe has been technically enabled to receive and process agentic transactions. What's missing is scale, which will come when the big AI platforms open agentic shopping to European consumers. Which Mellou referred to as "imminent", while warning against getting ahead of ourselves: "This will not happen overnight. We are still in a stage where we execute controlled pilots, we gather learnings from the ecosystem, and we are adapting before scaling.”

"Hey Roza, I wasn't able to fulfil your ask”

Before the trust panel, two more PIF members took the stage. Chris Ford, Senior Director, BHN EMEA, looked at how stored value can stay spendable in the age of agentic AI. Nail Osmanov, Digital Products UK&I at Visa, shared Visa's lessons from taking agentic commerce "from whiteboard to reality", and where the industry will need to work together next.

Andrew O'Connor, Agentic AI Lead at PSE Consulting, then brought Mellou, Zaman, BHN's Chris Ford and Visa's Nail Osmanov together for a panel on trust. The issuer view came through clearly. Customers need to be in control of their agents, "not the other way around". On regulation, Mellou argued the existing rules should "extend and adapt" rather than start again, and pointed to a year of momentum towards open protocols that stop the market fragmenting. Liability got an honest answer too: it's something the industry will need to road-test, and some of it may only be settled in court.

The clearest moment was Mellou's explanation of verifiable intent. She asked the room to imagine telling an agent to find her a pair of running trainers, within budget, and buy them next week. Then she goes to sleep.

If the agent finds exactly what she asked for, it goes ahead. If anything's off, because the trainers are out of stock or the price has gone up, the agent has to come back and say: "Hey Roza, I wasn't able to fulfil your ask." Then she authorises it again. It's a small amount of friction, she said, deliberately added "to bring the consumer back in". For now, Mastercard's pilots keep a human in the loop, with full delegated buying yet still to come.

Agentic-ready? Some of the room already is

The final session, introduced with new research from Edgar, Dunn & Company, saw Beatrice Sava put Leftwich, Stripe's Sean Werkema and Vaux on the spot. Leftwich was "very glad to go after Visa and Mastercard". Where they'd talked in months, he talked in days. The following week, he said, Dataiera would become the first platform globally running Mastercard Agent Pay and Visa Intelligent Commerce together in one channel. It starts with 8,000 employees in the UAE before going to the public a couple of weeks later. His merchant's product was a complicated one, with venues, seat maps and times, and the integration took a couple of days. He still warned against the hype that this is easy and repeatable.

Werkema said he had to rewrite his own remarks. When he started preparing, his line was that agentic commerce was "real in terms of infrastructure, but not really real yet in terms of transactions". In the few weeks since, that had "pretty dramatically changed". Volume used to be the lagging indicator, and now demand has caught up. The US is leading, he said, without the regulatory complications Europe has to work through.

Vaux said Thredd is seeing growth among crypto-friendly clients and in APAC and LATAM, with the UK and EU only now getting ready. He also had a call-out for the schemes: "It's not always helpful when Visa and Mastercard aren't completely aligned on this." Sava, speaking as a consultant, admitted she quite liked that, to laughter.

On which sectors go first, the panel didn't entirely agree. Leftwich backed travel, pointing out that almost every example Visa and Mastercard use is a travel one, with airlines and hotels sitting side by side in a single checkout. Vaux wasn't sure whether that's cause or effect: travel has always been full of technically advanced intermediaries, so of course it's early. He saw bigger potential in business travel, where company policies give an agent clear rules to follow.

Werkema believes that agents will win first where products are well defined and taste doesn't come into it: "If I say I want a blue-green sweater and it comes back with a green-blue sweater, that might lead to a reduction in consumer trust.”

Your daily coffee order is a safer bet, he said, as are services with fixed times, like a haircut or booking your dog into a spa. Stripe is also excited about agents paying other agents, with stablecoins making micropayments of a few cents per API call workable.

Vaux brought back his déjà vu. Many of today's concerns are "same old, same old", he said, no different from when contactless, wallets or Apple Pay arrived. Contactless took off once it became habitual and invisible. Apple Pay worked because it was only a small step on from tapping a card. Agentic commerce will probably grow the same way; from things people already do, with brands they already use every day.

From pilots to the mainstream

A few months ago, a day on agentic payments would have been mostly speculation. This one certainly wasn't. Mastercard's European issuers are technically enabled, Thredd is building controls for issuers, ANNA Money already runs most of its customer conversations through AI, and Dataiera was days from launch. 

There are still stones left unturned however, with trust still unsettled. Knowing which agent is acting, proving what the customer actually asked for, deciding who carries the risk when it goes wrong, and getting Visa and Mastercard pointing the same way. If Vaux is right, payments has been here before, with contactless and with wallets, and got there in the end. This time, the new face at the till just happens to be an AI agent.

Thank you to our speakers, to our sponsors Thredd and DB Seret, to our Lead Benefactor Mastercard and Benefactor BHN, to Edgar, Dunn & Company for their partnership, and to everyone who kept the conversation going into the evening.