Every time a Mastercard is tapped, the network has less than a tenth of a second to assess the likelihood of the purchase being fraudulent. Last year, 175 billion in transactions made this call. Now, the buyer on the other side of that decision is starting to change, and Greg Ulrich, the company’s chief artificial intelligence and data specialist, explained the result. VB Transform 2026 July 14 audience in Menlo Park. "We have established a number of risk rules to stop a bot from operating over time." Ulrich said. "Now we need to allow the bot to trade, which requires changes to our risk framework and risk rules."
Ulrich joined Mastercard eleven years ago when an analytics firm he worked for was acquired, and he says he was trusted from day one. "This allows a merchant who has never met you to accept payment and be sure they will get paid. This allows you as a consumer to transact and ensure that things are done in a safe, secure manner. If something goes wrong, there is a safe and secure way to dispute and resolve it," he said.
175 billion transactions were recorded in less than 100 milliseconds
He took the audience into each of these calls. "When you tap your Mastercard to pay for a product or service, we award points for that transaction." he said. "We have less than 100 milliseconds to look at it and give it a score from zero to 999 on how likely it is to be fake or real. And we pass it on to the issuing bank."
Generative AI has expanded what this account can see. "Because we have new technology, we can bring in more information, we can bring in more context, and now we’re seeing that we can identify 300, 400% more fraudulent transactions in these high-risk groups." Without adding friction or false positives for consumers, Ulrich said. of the company Safety net the system has stopped more than 70 billion fraudulent transactions, he told the audience, and Mastercard is building its transformative model on its transaction data as the foundation for new safety, security and personalization solutions. VentureBeat’s Beyond the pilot podcast earlier this year dissected that pile of manufacturing fraud in detail.
A third of service businesses are already powered by artificial intelligence
Business stakes reach past fraud. About 40% of Mastercard’s business is now services-based, Ulrich said, including marketing services; fraud, safety and security; and business intelligence. "A third of these are based on artificial intelligence, and they are growing faster than anything else." he said.
A line he returned to the entire session went further. "What will ensure that AI continues to scale is not the capabilities of agents, but how much we trust those agents on our behalf as consumers, businesses, financial institutions, or otherwise." he said.
Five layers stand between the agents and the network
Agent trading changes the protected object. "Instead of a single atomic operation where I say go buy something, I’m actually empowering or empowering the consumer, a business empowers," Ulrich said. "And when that happens, it’s a much more complicated operation." Trust, in turn, has a precondition. "The only way we can work with confidence is if we can define what the intent is, what the behaviors are, what the constraints are on this operation."
Ulrich went through the five layers that Mastercard has built against this problem. Personality comes first. "I want to make sure that I can understand not only who the consumer is, but who the agent is, that I put them together and that I have a KYA or that I know your agent, that I confirm that it’s a legitimate tech, that it’s a legitimate agent." he said. "We can register in our system."
Verifiable intent resolves "wrong – Nikes" problem
The verified intent is the second, an unfalsifiable cryptographic record of the original instructions that went with the transaction. "If you requested a size 12 Nike black Nikes but got them on sale and they’re not being returned and it’s not your order, here’s a way to look at it objectively and clearly on the back end." he explained.
Controls constitute the third layer, which determines which merchants the agent can buy from, with what limit, and under what restrictions. Implementation is underway Mastercard Agent Paybearing "tokenization, authentication, acceptance framework embedded in it" Ulrich said and started work with Microsoft, OpenAI, Google and others. Intelligence is the fifth layer, it covers the rules of risk, gives signs of insight "agreed or authorized access to concepts" personalized recommendations and monitoring via Recorded Future to identify threat actors in the system.
It is a purchasing agent with a larger premium budget
Consumer purchases are where agent trading begins. Ulrich pointed across the room from them to business-to-business buying as a bigger opportunity. His example was a manufacturer who wanted an always-on assembly line with an agent that manages inventory levels, monitors stock-outs, auto-replenishes, understands budget and approved suppliers. "When you start enabling this, the same five layers are required for this type of operation," he said.
Making it work between companies increases the parties that have to trust each other. "You need clear standards for identity, you need clear standards for intent, you need to work on those. You will have a purchasing agent, a supplier agent, a bank agent. For that to happen autonomously, they would all have to communicate, and that would require a really large-scale trust infrastructure."
Powerful new models, same safety action
Mastercard sat on the early wave Glasswing project Worked with Anthropic’s Mythos model and with OpenAI GPT-5.5-Cyberhe said. "What we’ve seen from both of these are incredibly powerful models that find new vulnerabilities in the ecosystem that were previously hard to detect, but it’s really a new tool as opposed to a new movement." Ulrich said.
Within the company, this work is led by the chief security officer. A dedicated team has prioritized the most critical assets, regularly traversing them between models, tracking findings by high, medium, and low severity, and using the same technology to manage patches. Ulrich said the approach has already been extended, and Mastercard is working to make the same architecture and patches available to others.
Mastercard will set it up differently in 14 months
"Guardrails, safety, all these things should be placed in the front part. These cannot be things we add on the back end. This is the first lesson. The second lesson is that you have to work for scale, and the other is as much about observability and accountability as intelligence." Ulrich said, recounting what the building inside Mastercard has taught the team. The company has built what he describes as an agent factory, with compliance, an observable operating system, and guardrails not attached to each agent. The model shift, which was once manually monitored by special teams, is now automated at that factory.
Asked by an audience member about earnings, Ulrich didn’t ease the trap from pilot to production. "If you are trying to extend it and then add guardrails as you extend it, once you’ve already built it, I think you’re doomed," he said.
Mastercard last year created a suite of agents for its 4,000 consultants, covering in-depth research, text to SQL, Excel and PowerPoint, at the expense of which Mastercard did not have the level it required. If the company started today, Ulrich said, it would have set them up completely differently. "I don’t know that fourteen months ago we would have expected to rethink the fundamental architecture and approach to building things."
Agent ID is connected to KYB and KYC
The personality layer is where Ulrich expects the market to move next. Within Agent Pay, Mastercard authenticates the consumer and connects the agent to that person, just like in traditional e-commerce. "Beyond that framework, I think there will be clear standards for defining who an agent is and connecting an agent with a consumer." he said. "And then we can tie that to a verifiable intent."
VentureBeat’s June 2026 Pulse research points in the same space. Only 32% of 107 qualified enterprise respondents give each agent its own comprehensive, manageable identityand only 12% include an agent-identification product in their consideration set.
He called the personality "one of the fastest growing ecosystems," Mastercard has been expanding there organically and inorganically for about six to seven years, he said, noting that the case is now covered. "agent identity as well as traditional KYB and KYC identity." The risk rules that keep bots off the grid come from more than two decades of applying AI to these operations. The rewrite for agents that Mastercard now wants to allow is already underway on the same network that collected 175 billion from them last year.
