Each time a Mastercard will get tapped, the community has lower than a tenth of a second to evaluate how seemingly the acquisition is to be fraudulent. It made that decision throughout 175 billion transactions final yr. Now the customer on the opposite facet of that judgment is beginning to change, and Greg Ulrich, the corporate's chief AI and knowledge officer, spelled out the consequence for the VB Remodel 2026 viewers in Menlo Park on July 14. "We've constructed a bunch of threat guidelines over time that have been meant to cease a bot from transacting," Ulrich stated. "Now we have to allow the bot to transact, in order that requires a change to our threat framework and our threat guidelines."
Ulrich joined Mastercard eleven years in the past when an analytics firm he labored at was acquired, and stated belief struck him from day one on the job. "It's what allows a service provider that's by no means met you to simply accept cost and make sure that they're going to receives a commission. It's what allows you as a client to transact and make sure that issues are going to work out in a trusted, safe means. And if one thing goes improper, there's a secure and safe path for a dispute and to resolve this," he stated.
175 billion transactions, scored in underneath 100 milliseconds
He took the viewers inside every of these calls. "Whenever you faucet your Mastercard to pay for a services or products, we're offering a rating to that transaction," he stated. "We have now underneath 100 milliseconds to take a look at that and provides a rating from zero to 999 about how seemingly is that to be fraudulent or actual. And we cross that on to the issuing financial institution."
Generative AI widened what that rating can see. "As a result of we have now new know-how, we will carry in additional knowledge, we will carry in additional context, and now we're discovering that we will determine 300, 400% extra fraudulent transactions at these high-risk bands," Ulrich stated, with out including friction or false positives for shoppers. The corporate's Security Web system has stopped greater than 70 billion fraudulent transactions, he instructed the viewers, and Mastercard is constructing its personal transformer mannequin on its transaction knowledge as a basis for brand spanking new security, safety, and personalization options. VentureBeat's Past the Pilot podcast took that manufacturing fraud stack aside intimately earlier this yr.
A 3rd of the companies enterprise already runs on AI
The enterprise stakes attain previous fraud. About 40% of Mastercard's firm is now primarily based on companies, Ulrich stated, together with advertising and marketing companies; fraud, security and safety; and enterprise intelligence. "A 3rd of these are predicated on AI, and people are rising at a a lot sooner clip than every little thing else," he stated.
One line he returned to all session went additional. "What's going to allow AI to proceed to scale shouldn’t be the capabilities of the brokers, it's how a lot we belief these brokers to do on our behalf as a client, as a enterprise, as a monetary establishment, or in any other case," he stated.
5 layers stand between brokers and the community
Agentic commerce modifications the article being secured. "As a substitute of a single atomic transaction the place I say go purchase one thing, I'm successfully delegating authority, or a client's delegating authority, a enterprise is delegating authority," Ulrich stated. "And when that occurs, it's a way more difficult transaction." Belief, in flip, has a precondition. "The one means it's going to work with belief is that if we will determine what was the intent, what are the behaviors, what are the constraints that have been meant in that transaction."
Ulrich walked by 5 layers Mastercard has constructed in opposition to that drawback. Identification comes first. "I wish to be sure that I can perceive not simply who the buyer is, however who the agent is, that I mix them collectively and that I’ve KYA or know your agent, that I'm validating that it's legit know-how, that it's a legit agent," he stated. "We will register it into our system."
Verifiable intent settles the "wrong-Nikes" drawback
Verifiable intent is second, a tamper-proof cryptographic file of the unique directions that travels with the transaction. "In case you've requested for Nike black Nikes in dimension 12, however you bought them on a closing sale they usually're not returnable and that wasn't in your instruction, there's a means to take a look at that in an goal and clear means on the again finish," he defined.
Controls kind the third layer, defining which retailers an agent should purchase from, at what restrict, and underneath what constraints. Execution runs by Mastercard Agent Pay, which carries "the tokenization, authentication, the acceptance framework embedded inside it" and has launched with Microsoft, OpenAI, Google, and others, Ulrich stated. Intelligence is the fifth layer, spanning threat guidelines, perception tokens that grant "consented or permissioned entry to insights" for customized suggestions, and monitoring by Recorded Future to determine risk actors within the system.
The larger prize is a procurement agent with a price range
Shopper purchases are the place agentic commerce began. Ulrich pointed the room previous them, to business-to-business procurement because the bigger alternative. His instance was a producer that wishes an always-on meeting line, with an agent that manages stock ranges, tracks when inventory runs low, replenishes robotically, and understands the price range and the permitted suppliers. "When you can begin enabling that, you require those self same 5 layers for that kind of transaction," he stated.
Making it work throughout corporations multiplies the events that should belief one another. "You want clear requirements for id, you want clear requirements for intent, you want these to work throughout. You're gonna have a procurement agent, a provider agent, a banking agent. They're all gonna want to speak to allow this to occur in an autonomous means, and that's gonna require actually scaled belief infrastructure."
Highly effective new fashions, identical safety movement
Mastercard sat within the early wave of Mission Glasswing with Anthropic's Mythos mannequin, and labored with OpenAI's GPT-5.5-Cyber, he stated. "What we've seen from each of these is extremely highly effective fashions discovering new vulnerabilities within the ecosystem that have been tough to detect beforehand, nevertheless it's actually a brand new instrument versus a brand new movement," Ulrich stated.
Inside the corporate, the chief safety officer leads that work. A devoted staff has prioritized probably the most essential belongings, runs them by the fashions routinely, tracks findings by excessive, medium, and low severity, and makes use of the identical know-how to deal with patches. Ulrich stated the method has already been prolonged out, and that Mastercard is working to make the identical structure and patching out there to others as nicely.
What Mastercard would construct in another way after 14 months
"The guardrails, the safety, all these items must be embedded on the entrance finish. These can't be issues that we're including on on the again finish. That's lesson one. Lesson two is you need to be working for scale, and the opposite one is round observability and accountability matter as a lot because the intelligence," Ulrich stated, counting off what constructing inside Mastercard taught the staff. The corporate constructed what he described as an agentic manufacturing unit, an working system with the compliance, the observability, and the guardrails in-built somewhat than bolted on per agent. Mannequin drift, as soon as tracked manually by devoted groups, is now automated into that manufacturing unit.
Requested by an viewers member in regards to the gotchas, Ulrich didn’t soften the pilot-to-production lure. "In case you're attempting to increase that after which add guardrails in as you're extending it, when you've already constructed it, I feel you're doomed to fail," he stated.
Mastercard constructed a collection of brokers final yr for its 4,000 consultants, overlaying deep analysis, textual content to SQL, Excel, and PowerPoint, instruments that by his account didn’t exist on the degree Mastercard wanted. Had been the corporate beginning right this moment, Ulrich stated, it will construct them basically in another way. "I don't know that we anticipated after we constructed issues fourteen months in the past that we might be rethinking the basic structure and the method already."
Agentic id joins KYB and KYC
The id layer is the place Ulrich expects the market to maneuver subsequent. Inside Agent Pay, Mastercard authenticates the buyer the best way it does in conventional e-commerce and binds the agent to that particular person. "Exterior of that framework, I feel there can be open requirements to determine who an agent is and bind the agent with the buyer," he stated. "After which we will tie that with verifiable intent."
VentureBeat's June 2026 Pulse analysis factors on the identical hole. Solely 32% of the 107 certified enterprise respondents give each agent its personal scoped, managed id, and simply 12% embody an agent-identity product of their consideration set.
He referred to as id "one of many faster-growing ecosystems," noting Mastercard has been increasing there organically and inorganically for about six or seven years, with the work now spanning "agentic id in addition to the normal KYB and KYC id." The chance guidelines that maintain bots off the community got here out of greater than twenty years of making use of AI to these transactions. The rewrite, for the brokers Mastercard now desires to let in, is already underway on the identical community that scored 175 billion of them final yr.
