Each time a Mastercard is tapped, the system has less than a tenth of a second to assess the likelihood of fraud in the transaction. Last year, this analysis was performed on 175 billion transactions. The profile of the purchaser subject to this assessment is evolving, as explained by Greg Ulrich, Mastercard’s chief AI and data officer, at the VB Transform 2026 event in Menlo Park on July 14. “We’ve crafted numerous risk rules over time to prevent bots from transacting,” Ulrich noted. “Now, we need to enable bots to transact, necessitating a shift in our risk structure and rules.”
Ulrich became part of Mastercard eleven years ago through an acquisition of his previous analytics firm. From the start, he was impressed by the company’s emphasis on trust. “Trust is what allows a merchant who hasn’t met you to accept payment, ensuring they will be compensated. It’s what empowers you as a consumer to make transactions with the assurance of security and resolution options,” he remarked.
175 billion transactions, scored in under 100 milliseconds
Ulrich elaborated on the rapid decision-making process involved in each transaction. “Upon tapping your Mastercard for a purchase, we assign a score to that transaction,” he stated. “We have under 100 milliseconds to analyze and provide a score from zero to 999 indicating the probability of fraud, which we then forward to the issuing bank.”
Generative AI has expanded the scope of this scoring. “With new technology, we can incorporate more data and context, enabling us to identify 300 to 400% more fraudulent transactions in high-risk categories,” Ulrich explained, all without increasing consumer inconvenience or false positives. Mastercard’s Safety Net system has blocked over 70 billion fraudulent transactions, and the company is developing its own transformer model based on transaction data to enhance safety, security, and personalization. VentureBeat’s Beyond the Pilot podcast has detailed this fraud prevention framework earlier this year.
A third of the services business already runs on AI
The implications extend beyond fraud prevention. Ulrich mentioned that about 40% of Mastercard is now service-based, encompassing marketing, fraud prevention, safety, security, and business intelligence. “A third of these services rely on AI and are expanding more rapidly than other sectors,” he remarked.
He emphasized, “The key to AI’s continued growth is not the agents’ capabilities but the trust we place in them to act on behalf of consumers, businesses, financial institutions, and others.”
Five layers stand between agents and the network
Agentic commerce alters the nature of secured transactions. “Instead of a single transaction where I authorize a purchase, authority is delegated, making it a more intricate transaction,” Ulrich stated. “Trust requires understanding the intent, behaviors, and constraints intended in the transaction.”
Ulrich described the five layers Mastercard has implemented to address these challenges. The first is identity. “We need to comprehend not just the consumer’s identity but also the agent’s, combining them and implementing KYA or ‘know your agent,’ ensuring it’s legitimate technology,” he said. “We can then register it into our system.”
Verifiable intent settles the “wrong-Nikes” problem
Verifiable intent forms the second layer, providing a cryptographic record of the original instructions accompanying the transaction. “If you requested black Nike shoes in size 12, but received them as a non-returnable final sale and that wasn’t specified, there’s a way to objectively verify that on the backend,” he explained.
The third layer, controls, defines which merchants an agent can transact with, specifying limits and constraints. Execution occurs via Mastercard Agent Pay, which incorporates tokenization, authentication, and an acceptance framework, launched with partners like Microsoft, OpenAI, and Google, Ulrich mentioned. The fifth layer, intelligence, includes risk rules, insight tokens for personalized recommendations, and monitoring to identify threats through Recorded Future.
The bigger prize is a procurement agent with a budget
Agentic commerce began with consumer purchases, but Ulrich sees greater potential in business-to-business procurement. He illustrated with a manufacturer seeking an always-operational assembly line, managed by an agent overseeing inventory, automatic restocking, budgeting, and supplier approvals. “Enabling this requires the same five layers,” he said.
Expanding this across companies increases the number of parties needing mutual trust. “Clear standards for identity, intent, and communication among procurement, supplier, and banking agents are essential for autonomous functioning, demanding a robust trust infrastructure,” he noted.
Powerful new models, same security motion
Mastercard has engaged with Anthropic’s Mythos model through Project Glasswing and collaborated with OpenAI’s GPT-5.5-Cyber, as Ulrich shared. “Both models have uncovered previously undetected vulnerabilities, serving as new tools rather than new approaches,” he said.
The company’s chief security officer oversees this work, prioritizing critical assets, running them through the models, categorizing findings by severity, and using the same technology for patch management. Ulrich mentioned that Mastercard aims to extend this architecture and patching capability to others.
What Mastercard would build differently after 14 months
Reflecting on Mastercard’s experience, Ulrich stressed, “Security must be integrated from the start, not added later. Scale is crucial, and accountability and observability are as important as intelligence.” He described the development of an agentic factory, an operating system with built-in compliance, observability, and guardrails rather than per-agent additions. Model drift, once manually tracked, is now automated.
In response to an inquiry about potential pitfalls, Ulrich cautioned against adding guardrails after development. “Attempting to retrofit safeguards is a recipe for failure,” he said.
Mastercard’s recent development of agents for its 4,000 consultants, which included tools for research, text to SQL, Excel, and PowerPoint, highlighted a need for fundamentally different approaches. Ulrich admitted, “We didn’t foresee rethinking our architecture and approach so soon after initial development.”
Agentic identity joins KYB and KYC
Ulrich anticipates the next market shift in the identity layer. Within Agent Pay, Mastercard authenticates consumers as in traditional e-commerce and links agents to them. “Outside this framework, open standards for agent identification and consumer binding are likely,” he said. “This can be combined with verifiable intent.”
VentureBeat’s June 2026 Pulse research highlights a gap, with only 32% of 107 enterprises providing each agent with a managed identity, and just 12% considering an agent-identity product.
Identity is “a rapidly growing ecosystem,” Ulrich noted, with Mastercard expanding in this area for six to seven years, covering agentic identity alongside traditional KYB and KYC identity. The risk rules developed over decades to exclude bots from the network are being revised to accommodate agents, utilizing the same infrastructure that processed 175 billion transactions last year.
